Abrasion monitoring method for carbon fiber reinforced composite drill bit
Through a multi-signal fusion monitoring method based on current signal-based model driving and deep neural network, the problems of waste of resources and low efficiency in the existing drill bit wear detection methods are solved, and accurate real-time monitoring of drill bit wear is realized, and processing quality is improved. efficiency.
Patent Information
- Application Number
- CN202410906428.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-08
- Publication Date
- 2024-12-06
AI Technical Summary
The existing drill bit wear detection methods have problems such as waste of resources resulting in premature tool replacement and failure to replace wear tools in time to affect processing quality. Offline observation reduces processing efficiency and increases production risks.
Using a model-based model to drive the drill bit wear monitoring model, combined with deep neural network and improved Kalman filtering algorithm, the drill bit wear is monitored through multi-signal fusion, adjust the weights in real time, extract the time and frequency domain characteristics, and achieve accurate drill bit wear. monitor.
It realizes the timely detection of tool failure problems, ensures production and processing quality, improves processing efficiency, reduces environmental noise interference, and balances the accuracy and generalization capabilities of model-driven and data-driven methods.
Smart Images

Figure CN119077854A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of drill bit wear monitoring, and in particular to a method for monitoring the wear of a carbon fiber reinforced composite material drill bit. Background Art
[0002] In the aerospace field, carbon fiber reinforced composite materials are widely used in the construction of aerospace bodies due to their high strength, light weight, corrosion resistance and other advantages. The drilling quality of carbon fiber reinforced composite materials has a significant impact on the overall assembly quality of the body and the safety of the product in later use. During the drilling process, drill wear will reduce the stability of the drilling process, and in severe cases, it will cause the drill to break, thereby affecting the drilling quality and processing efficiency. Therefore, the prediction of drilling force and the monitoring of drill wear during the drilling process are particularly important.
[0003] The main problems of existing drill wear detection methods include:
[0004] At present, the industry mainly uses empirical life analysis and offline observation methods to monitor drill bit status and quality. The empirical life analysis method is based on the empirical life of the tool. The used tool will be replaced when it reaches the empirical life value, but this will bring two problems:
[0005] (1) Replacing the tool too early will result in the tool not being fully used within its limited lifespan. Failure to replace worn tools in a timely manner will make it difficult to ensure processing quality, resulting in a decrease in yield rate.
[0006] (2) The offline observation method reduces the processing efficiency and cannot detect tool failure problems in the first place.
[0007] In summary, these two methods reduce processing efficiency while increasing production risks. Summary of the invention
[0008] The object of the present invention is to provide a method for monitoring wear of a carbon fiber reinforced composite material drill bit to overcome the above-mentioned defects in the prior art.
[0009] A method for monitoring wear of a carbon fiber reinforced composite material drill bit comprises the following steps:
[0010] S1. Considering tool wear, a model-driven drill wear monitoring model based on current signal is established in combination with the electric spindle current sensor signal, which includes: derivation of drill main cutting edge load model, derivation of drill chisel edge load model, derivation of drilling model load considering drill wear, and modeling of drilling load-spindle current model;
[0011] S2. Analyze the wear process of the drill bit to obtain the wear characteristics of the drill bit, and select multiple sensor signals as monitoring signals of the drill bit wear status in combination with the drilling environment and conditions, and perform feature extraction of the sensor signals in the time domain and frequency domain;
[0012] S3. Using the extracted characteristic signals, a data-driven multi-signal fusion drill wear monitoring model based on deep neural network is established;
[0013] S4. Combined with the model driven drill wear monitoring model based on current signal, a model based on improved Kalman filter algorithm and a data fusion driven drill wear monitoring model are established to realize the monitoring of drill wear during the drilling process.
[0014] Preferably, in step S1, considering that it is difficult to establish an accurate analytical model for carbon fiber reinforced composite materials, a semi-empirical formula is used to model the drilling process:
[0015]
[0016] In the formula, F vu and F hu are the forces on the cutting edge unit in the vertical and horizontal directions, respectively; A and B are parameters related to drilling parameters and drill geometry to be determined experimentally; γ is the drill rake angle; and h is the cutting depth.
[0017] Preferably, in the derivation of the drill bit main cutting edge load model in step S1, when the drill bit is drilling carbon fiber reinforced composite material, the main cutting edge of the drill bit completely cuts into the workpiece material, which can be regarded as a series of adjusted unit orthogonal cutting, and the main cutting edge is divided into a number of cutting units according to the mechanical analysis method and the microelement method;
[0018] At any cutting position angle θ, the cutting layer thickness h and the cross-sectional area of the cutting unit are:
[0019]
[0020] Where φ is half of the drill bit angle, f r The feed rate per revolution of the drill is:
[0021]
[0022] Where n is the spindle speed (krpm), f is the spindle feed speed (mm / s). For the convenience of calculation, the coordinate system is set to the polar coordinate system, and the radial distance is normalized:
[0023]
[0024] Where ρ is the radial coordinate in the polar coordinate system, θ(ρ) is the cutting edge inclination angle at the radial coordinate ρ, R is the drill radius, r is the distance from a point on the main cutting edge to the center, δr is the cutting edge unit width at the radial coordinate ρ, and the cosine value can be obtained from the drill geometry:
[0025]
[0026] Where b is the drill core thickness. Combined with the above formula, the force F generated by the entire cutting edge perpendicular to the cutting edge direction and the cutting speed direction is v And the cutting speed direction force F h It can be obtained by integrating the unit force:
[0027]
[0028] In the formula, the entire cutting edge is calculated, the upper limit of the integral is: j = R / R = 1, the lower limit of the integral is: i = (b / sinψ) / R, ψ is the chisel edge angle, and the axial force F T is the force F perpendicular to the two cutting edges v Axial force F a The resultant force, drilling torque M h is the integral of the moment acting on the cutting element, as follows:
[0029]
[0030] In the formula, the drill bit rake angle γ is divided into two parts: the drill bit rake angle itself - static rake angle γ s The second part is determined by the feed rate - dynamic rake angle γ d , as shown below:
[0031]
[0032] Preferably, the micro-element force on the chisel edge in the derivation of the drill chisel edge load model in step S1 can be expressed as:
[0033]
[0034] Where E3 is the Young's modulus of the workpiece material in the feed direction, v is the Poisson's ratio of the workpiece material, and γ ω is the half angle of the chisel edge, γ d It is the angle between the total cutting force of the chisel edge and the vertical direction, that is, the aforementioned dynamic rake angle. Due to the existence of the dynamic rake angle, the axial force on the chisel edge can be expressed as:
[0035]
[0036] Preferably, the cutting force caused by tool wear in the drilling model load derivation considering drill wear in step S1 can be divided into two parts: F perpendicular to the tool wear area and F perpendicular to the tool wear area. vw and the friction force F parallel to the wear area hw , whose expression is as follows:
[0037]
[0038] Where H is the hardness of the material, VB is the width of the single-edge wear zone, s is the length of the wear zone, μ is the friction coefficient between the workpiece and the tool material, and the axial force F generated by drill wear is Tw F vw The axial force F aw Assuming that the two edges of the drill are evenly worn during drilling, the torque M generated by the wear of the drill flank is w The differential element method and double integral calculation are used in the wear area as follows:
[0039]
[0040] Where α is the back angle of the main cutting edge, l is the projection length of the main cutting edge on the xy plane, S w is the projection of the drill bit back face wear area on the xy plane. According to the above formula, the tool wear factor can be added to modify the formula of the axial force and torque of the drill bit, as shown in formula (1-13). According to this formula, the axial force and torque of the drill bit under different wear degrees can be calculated;
[0041]
[0042] Preferably, in the drilling load-spindle current modeling of step S1, the spindle current and the torque on the drill bit have the following relationship model:
[0043]
[0044] In the formula, K t is the torque constant of the spindle motor, I is the load current of the spindle motor, T m is the load torque acting on the spindle motor, T f is the internal friction torque overcome when the spindle motor rotates, J is the equivalent moment of inertia of the spindle, ω is the angular frequency of the spindle, M total is the torque on the drill bit, see formula (1-13);
[0045] In the actual drilling process, under the control of the electric spindle driver, the rotation speed of the spindle motor remains unchanged, that is, the spindle angular acceleration is 0 (dω / dt=0), so the equivalent rotational inertia J of the spindle can be ignored. When the spindle motor is idling, it only needs to overcome the internal friction torque of the motor, and the torque on the drill bit is 0 (M total =0), then:
[0046] K t I0(n)=T f (1-15)
[0047] In the formula, I0(n) is the idling current when the spindle motor speed is n. Substituting formula (1-15) into formula (1-14), the drilling current increment △I is obtained as follows:
[0048]
[0049] It can be seen from the formula that the spindle current increment is proportional to the load on the drill bit, thus establishing a theoretical model of drilling load and spindle current. Combined with the drill wear-drilling load model, the drill wear status can be monitored by collecting the real-time spindle current of the electric spindle during drilling.
[0050] Substituting equations (1-7), (1-12), and (1-13) into (1-16), we obtain:
[0051]
[0052] From this formula, it can be seen that the drilling current increment △I of the high-speed spindle can be expressed by the drilling parameters (spindle speed n and feed speed f) and the single-edge wear band width VB, that is, △I=f(n,f,VB). A, B, μ, H, K in formula (1-17) and (1-13) are t , E3, v and other parameters are constants related to the drill geometry parameters, drilling parameters, workpiece material properties and electric spindle characteristics, and can be determined by fitting through variable parameter drilling load and spindle current measurement experiments.
[0053] Preferably, in step S2, current signal feature extraction: the current signal of the drill bit when working is collected by a high-speed data acquisition card, and 15 time domain features are extracted; vibration signal feature extraction: 2 miniature three-axis acceleration sensors are placed on the spindle fixture and the workpiece in combination with the NI signal acquisition system to collect vibration signals, extract time domain and focus on extracting frequency domain features, a total of 229 features; acoustic emission signal feature extraction uses a SAEU2S acoustic emission system to collect acoustic emission signals, the sampling frequency is 1 MHz, and the acoustic emission signals of drill bits with different degrees of wear during the drilling process are collected, 19 time domain features are extracted, and a total of 263 features are extracted from the current signal, vibration signal and acoustic emission signal.
[0054] Preferably, in step S3, the extracted features are normalized to produce feature samples as input of the deep neural network, and the drill wear value is used as output for model training to establish a data-driven monitoring model based on the deep neural network.
[0055] Preferably, in step S4, data-driven multi-signal fusion monitoring of drill bit wear is realized through a deep neural network model (DNN), Dropout and model parameter optimization are used to reduce the root mean square error of the model and improve the accuracy of the model, and then an improved Kalman filter algorithm is used to realize model and data fusion driven monitoring of drill bit wear.
[0056] The present invention has the following advantages:
[0057] 1. The present invention can adjust the weight online in real time, can promptly discover tool failure problems, ensure production and processing quality, and improve processing efficiency.
[0058] 2. The present invention greatly improves the interference of environmental noise on data by extracting the characteristics of sensor signal data.
[0059] 3. Although the traditional model-driven drill wear monitoring method is more accurate than the data-driven drill wear detection method, its generalization ability is average. The present invention balances the advantages and disadvantages of the two methods and improves the generalization ability of the drill wear monitoring method while ensuring sufficient accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a schematic diagram of the force on the main cutting edge of the carbon fiber reinforced composite material drill bit of the present invention when drilling.
[0061] Figure 2 The present invention is a schematic diagram of the force applied to the chisel edge of the carbon fiber reinforced composite material drill bit during drilling.
[0062] Figure 3 It is a schematic diagram of the additional force generated by the wear of the flank of the drill bit of the present invention.
[0063] Figure 4 It is a schematic diagram of the total force of the drill bit of the present invention during drilling.
[0064] Figure 5 It is a flow chart of drill bit wear status monitoring based on drilling current of the present invention.
[0065] Figure 6 It is the original acoustic emission signal collected by the present invention.
[0066] Figure 7 The invention relates to a multi-sensor signal acquisition device.
[0067] Figure 8 It is a schematic diagram of the deep neural network structure of the present invention.
[0068] Fig. 9 It is the standard neural network of the present invention and the neural network after adding Dropout.
[0069] Fig.10 It is the hidden layer (3 layers) deep neural network structure of the present invention.
[0070] Fig.11 It is the deep neural network model structure after parameter adjustment of the present invention.
[0071] Fig.12 It is a schematic diagram of the model and data fusion driven drill wear monitoring model of the present invention.
[0072] Fig.13 It is a schematic diagram of the model based on improved Kalman filtering and data fusion driven modeling of the present invention. DETAILED DESCRIPTION
[0073] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the concept and technical solution of the present invention.
[0074] A method for monitoring wear of a carbon fiber reinforced composite material drill bit comprises the following steps:
[0075] S1. Considering tool wear, a model-driven drill wear monitoring model based on current signal is established.
[0076] S2. The drill wear process is analyzed to obtain the wear characteristics of the drill bit. Combined with the drilling environment and conditions, multiple sensor signals are selected as monitoring signals of the drill wear status, and the sensor signals are feature extracted.
[0077] S3. Using the extracted feature signals, a data-driven multi-signal fusion drill wear monitoring model based on deep neural network is established.
[0078] S4. Combined with the model driven drill wear monitoring model based on current signal, a model based on improved Kalman filter algorithm and a data fusion driven drill wear monitoring model are established to realize the monitoring of drill wear during the drilling process.
[0079] 1. In step S1, a model driven drill wear monitoring model based on current signals is established.
[0080] At present, the drilling force modeling methods of composite materials mainly include empirical method, analytical method, finite element method, etc. Considering that it is difficult to establish an accurate analytical model of carbon fiber reinforced composite materials, a semi-empirical formula is used to model the drilling process.
[0081]
[0082] In the formula, F vu and F hu are the forces on the cutting edge unit in the vertical and horizontal directions, respectively; A and B are parameters related to drilling parameters and drill geometry to be determined experimentally; γ is the drill rake angle; and h is the cutting depth.
[0083] The analytical modeling of carbon fiber reinforced composite drilling load mainly focuses on the loads on the main cutting edge and chisel edge of the drill during the drilling process and the additional loads generated on the drill after the drill is worn.
[0084] 1.1 Derivation of drill bit main cutting edge load model
[0085] The commonly used carbon fiber reinforced composite drill is a double-edged drill. When the main cutting edge of the drill is fully cut into the workpiece material when drilling carbon fiber reinforced composite materials, it can be regarded as a series of adjusted unit orthogonal cutting. According to the mechanical analysis method and the microelement method, the main cutting edge is divided into several cutting units, such as Figure 1 shown.
[0086] At any cutting position angle θ, the cutting layer thickness h and the cross-sectional area of the cutting unit are:
[0087]
[0088] Where φ is half of the drill bit angle, f r The feed rate per revolution of the drill is:
[0089]
[0090] Where n is the spindle speed (krpm), and f is the spindle feed speed (mm / s). For the convenience of calculation, the coordinate system is set to the polar coordinate system, and the radial distance is normalized:
[0091]
[0092] Where ρ is the radial coordinate in the polar coordinate system, θ(ρ) is the cutting edge inclination angle at the radial coordinate ρ, R is the drill radius, r is the distance from a point on the main cutting edge to the center, δr is the cutting edge unit width at the radial coordinate ρ, and the cosine value can be obtained from the drill geometry:
[0093]
[0094] In the formula, b is the drill core thickness. Combined with the above formula, the force F generated by the entire cutting edge perpendicular to the cutting edge direction and the cutting speed direction is v And the cutting speed direction force F h It can be obtained by integrating the unit force:
[0095]
[0096] In the formula, the entire cutting edge is calculated, the upper limit of the integral is: j = R / R = 1, the lower limit of the integral is: i = (b / sinψ) / R, ψ is the chisel edge angle. Axial force F T is the force F perpendicular to the two cutting edges v Axial force Fa The resultant force, drilling torque M h is the integral of the moment acting on the cutting element, as follows:
[0097]
[0098] In the formula, the drill bit rake angle γ is divided into two parts: the drill bit rake angle itself - static rake angle γ s The second part is determined by the feed rate - dynamic rake angle γ d , as shown below:
[0099]
[0100] 1.2 Derivation of drill bit chisel edge load model
[0101] When drilling composite materials, studies have shown that the chisel edge of the drill has a great influence on the axial force. The drilling force generated by the chisel edge accounts for about 60% of the total drilling force, and the chisel edge mainly has an extrusion effect on the material rather than a cutting effect. Figure 2 As shown. The micro-element force on the chisel edge can be expressed as:
[0102]
[0103] Where E3 is the Young's modulus of the workpiece material in the feed direction, v is the Poisson's ratio of the workpiece material, and γ ω is the half angle of the chisel edge, γ d It is the angle between the total cutting force of the chisel edge and the vertical direction, which is the aforementioned dynamic rake angle. Due to the existence of the dynamic rake angle, the axial force on the chisel edge can be expressed as:
[0104]
[0105] Since the chisel edge mainly exerts an extrusion effect on the material and has little wear, the change in drilling torque caused by chisel edge wear is ignored in this study.
[0106] 1.3 Derivation of drilling model load considering drill wear
[0107] In the process of composite material drilling force modeling, many scholars have established composite material drilling force models based on empirical and analytical methods. However, previous studies only considered the influence of tool geometry parameters, drilling parameters and composite material properties, but did not consider the influence of drill bit wear on drilling load. This section will model the additional drilling load caused by drill bit wear.
[0108] The wear of drilling drill bits is mainly manifested as progressive triangular wear of the main cutting edge and the back face of the chisel edge. The cutting force caused by tool wear can be divided into two parts: F perpendicular to the tool wear area vw and the friction force F parallel to the wear areahw ,like Figure 3 Its expression is as follows:
[0109]
[0110] Where H is the hardness of the material (Brinell hardness value), VB is the width of the single-edge wear zone, s is the length of the wear zone, and μ is the friction coefficient between the workpiece and the tool material. Figure 3 As shown, the axial force F generated by drill wear Tw F vw The axial force F aw Assuming that the two edges of the drill are worn evenly during drilling, the torque M generated by the wear of the drill flank is w The differential element method and double integral calculation are used in the wear area as follows:
[0111]
[0112] Where α is the back angle of the main cutting edge, l is the projection length of the main cutting edge on the xy plane, S w is the projection of the drill bit back face wear area on the xy plane. According to the above formula, the tool wear factor can be added to modify the formula of the axial force and torque of the drill bit, as shown in formula (1-13). The axial force and torque of the drill bit under different wear degrees can be calculated according to this formula. The total drilling force and torque of the drill bit are as follows: Figure 4 As shown in the figure, the model extends the drilling load model to the entire cutting cycle, which increases the practicality of the model.
[0113]
[0114] 1.4 Drilling load-spindle current model
[0115] From the previous drilling wear-drilling load modeling, it can be seen that during drilling, as the drill wear increases, the drilling load changes, which is mainly manifested as an increase in the axial force and torque on the drill, which changes the load of the spindle motor and causes a change in the load current of the spindle motor. Since the axial drilling force is parallel to the spindle rotation center and does not generate torque on the drill, the load of the spindle motor is mainly caused by the torque acting on the drill and the friction torque inside the spindle. The spindle current and the torque on the drill have the following relationship model:
[0116]
[0117] In the formula, K t is the torque constant of the spindle motor, I is the load current of the spindle motor, T m is the load torque acting on the spindle motor, T fis the internal friction torque overcome when the spindle motor rotates, J is the equivalent moment of inertia of the spindle, ω is the angular frequency of the spindle, M total is the torque acting on the drill bit, see formula (1-13).
[0118] In the actual drilling process, under the control of the electric spindle driver, the rotation speed of the spindle motor remains unchanged, that is, the spindle angular acceleration is 0 (dω / dt=0), so the equivalent rotational inertia J of the spindle can be ignored. When the spindle motor is idling, it only needs to overcome the internal friction torque of the motor, and the torque on the drill bit is 0 (M total =0), then:
[0119] K t I0(n)=T f (1-15)
[0120] In the formula, I0(n) is the idling current when the spindle motor speed is n. Substituting formula (1-15) into formula (1-14), the drilling current increment △I is:
[0121]
[0122] It can be seen from the formula that the spindle current increment is proportional to the load on the drill bit, thus establishing a theoretical model of drilling load and spindle current. Combined with the drill wear-drilling load model, the drill wear status can be monitored by collecting the real-time spindle current of the electric spindle during drilling, such as Figure 5 shown.
[0123] Substituting equations (1-7), (1-12), and (1-13) into (1-16), we obtain:
[0124]
[0125] From this formula, it can be seen that the drilling current increment △I of the high-speed spindle can be expressed by the drilling parameters (spindle speed n and feed speed f) and the single-edge wear band width VB, that is, △I=f(n,f,VB). A, B, μ, H, K in formula (1-17) and (1-13) t , E3, v and other parameters are constants related to the drill geometry parameters, drilling parameters, workpiece material properties and electric spindle characteristics, and can be determined by fitting through variable parameter drilling load and spindle current measurement experiments.
[0126] 2. In step S2, the data collected by the sensor signal will be subject to many interferences from the surrounding environment, and the original signal is difficult to be used directly for wear monitoring. Therefore, it is necessary to extract features from the sensor signal through preprocessing and transformation to extract features with high reliability and high sensitivity in tool wear. Since the calculation efficiency in the time-frequency domain is not high and the extraction requires the use of algorithms such as empirical mode decomposition, only the features in the time domain and frequency domain are extracted. The time domain features directly calculate the statistical features of the time domain signal. The following Table 2-1 shows the time domain features and calculation formulas. The frequency domain features calculate the statistical features of the original signal after performing FFT (Fast Fourier Transform). The following Table 2-2 shows the frequency domain features and calculation formulas. Figure 6 It is the original acoustic emission signal collected during the drilling process of drills with different degrees of wear. The difference between the acoustic emission signals with different degrees of wear can be seen from the time domain waveform. In addition to the violent fluctuations of drilling in and out, the amplitude of the drill with greater wear is generally larger during the drilling stage. Extract time domain features. Since the frequency range of the acoustic emission signal is in the high frequency region (100kHz~1MHz), it is not easily affected by environmental noise, so the frequency domain features are extracted. A total of 19 time domain and frequency domain features are extracted.
[0127] Table 2-1 Time domain characteristics and calculation formula
[0128]
[0129] Table 2-1 Time domain characteristics and calculation formula
[0130]
[0131]
[0132] Table 2-2 Frequency domain characteristics and calculation formula
[0133]
[0134] 3. In step S3, if Figure 7A multi-signal acquisition system based on the synchronous acquisition of current signal, vibration signal and acoustic emission signal was built, and three identical double-edged carbide drills were used in the experiment to drill on a carbon fiber reinforced composite drilling machine. The workpiece material was 1.6mm thick carbon fiber reinforced composite material, the spindle speed n = 60krpm, the spindle feed speed was 20mm / s, and each drill was used to drill 4000 holes. Every 50 holes, the ultra-depth of field three-dimensional microscopy was used to shoot and measure the average wear VB on the two main cutting edges of the drill, and the current signal, vibration signal and acoustic emission signal were collected during the drilling process. The sampling frequency of the current signal and vibration signal was 50kHz, and the sampling frequency of the acoustic emission signal was 1MHz. 70% of the 243 experimental data of the three drills were randomly selected as the training data set, and 30% as the test data set.
[0135] The use of deep neural networks (DNN) in the field of machine learning further improves the model's capabilities. The structure of a deep neural network is as follows Figure 8 shown.
[0136] Deep neural networks can be divided into input layer, hidden layer and output layer according to the type of layers. The first and last layers are the input layer and output layer, and the intermediate layers are all hidden layers. Each hidden layer has multiple neurons, and the hidden layers are fully connected, that is, a neuron in a hidden layer is connected to any neuron in the next hidden layer. The basic idea of the deep neural network algorithm is composed of the forward propagation of the network and the back propagation of the error: forward propagation is the input vector through multiple weight coefficient matrices and offset vectors to perform linear operations and neuron activation function operations, and finally obtain the output layer from the input layer layer by layer, as shown below:
[0137] a l =σ(z l )=σ(W l a l-1 +b l )
[0138] In the formula, σ(z l ) is the activation function of the neuron, a l is the output of the hidden layer and the output layer, W l is the weight coefficient matrix of the lth layer, b l is the offset vector of the lth layer.
[0139] The loss function is used to measure the loss between the output value obtained by the forward calculation and the true value, as shown below:
[0140]
[0141] In the formula, ||||2 is the L2 norm, a Lis the output value of the output layer, and y is the true value. Back propagation uses batch gradient descent method (Batch) to optimize the loss function to find the minimum value. In this process, the weight value and threshold are continuously adjusted to obtain the optimal weight coefficient matrix and offset vector corresponding to each layer, and finally the output value is as close to or equal to the true value as possible:
[0142]
[0143] In the formula, α is the iteration step size, m is the input sample size, and δ i,l The output layer’s δ is calculated by the loss function gradient i,L After recursion layer by layer, we get:
[0144] δ i,l =(W l+1 ) T δ i,l+1 ⊙σ′(z i,l )
[0145] Where ⊙ is the Hadamard product. For two vectors of the same dimension A(a1, a2, …, a n ) T and B(b1,b2,…,b n ) T , then A⊙B=(a1b1,a2b2,…,a n b n ) T .
[0146] The activation function mentioned in the formula is introduced to enhance the nonlinear relationship expression ability of the neural network and avoid the multi-layer network being equivalent to a linear function. The activation function has the characteristics of nonlinearity, differentiability and monotonicity. The present invention studies the online monitoring of drill wear. The output of the model is the drill flank wear value VB for different drilling numbers, which belongs to a regression problem. The present invention selects the tanh function suitable for the regression model as the activation function of the hidden layer and the linear function as the activation function of the output layer.
[0147] Since deep neural networks contain multiple hidden layers, and each hidden layer has multiple neurons, when there are too many parameters, after model training, the model may have a good fitting effect on the training set, but may not accurately predict unknown data, which is called overfitting. In order to avoid overfitting in neural networks, this application chooses a relatively simple and easy-to-use dropout method, that is, in the hidden layer of the neural network, a fixed probability p is set, and each neuron is judged whether to keep it based on this probability, such as Fig. 9 shown.
[0148] For a neural network y = f(wx + b), introduce a discard function d() so that y = f(wd(x) + b). The discard function is defined as:
[0149]
[0150] When the model is in the training phase, m∈{0,1} is the discard mask, which is randomly generated by probability p, and p is generally taken as 0.5. Assuming that there are 10 neurons in a hidden layer, let p be 0.5, and the randomly generated discard mask m={0,0,1,0,1,1,1,0,1,0}, the 1st, 2nd, 4th, 8th, and 10th neurons of this layer will be discarded during the training phase. In the test phase, neurons are not discarded, all neurons are retained, and the output of each neuron is multiplied by probability p, which is equivalent to averaging different neural networks to ensure the consistency of output during the training and test phases. Under such operations, the mutual dependence between neurons can be reduced, overfitting can be reduced, and the accuracy of the model can be improved.
[0151] In order to evaluate the performance of the deep neural network model and quantify the model accuracy, the root mean square error RMSE and the mean absolute percentage error MAPE are used as the evaluation indicators of the model, as shown in the following formula:
[0152]
[0153] In the formula, n represents the number of samples, Y i is the i-th sample value, is the predicted value corresponding to the model.
[0154] The multi-sensor signal acquisition system collects vibration, acoustic emission and current signals of the processing process through sensors, and then extracts characteristic values after data cleaning and signal processing. This study extracted 263 features based on the processing technology and signal characteristics. The reason why the features are not reduced in dimension and screened here is that the neural network itself has the ability to screen features, and the data set capacity is not large and the running time is not long, so the neural network is used to screen the features, and the contribution of each feature to the prediction model can be observed.
[0155] This application builds a multi-signal fusion drill wear status monitoring model based on deep neural network based on Python language, TensorFlow machine learning library and Keras machine learning interface. First, the data set is divided into training set and test set with a ratio of 7:3. Considering the small data set, a neural network classification model with 3 hidden layers will be designed. The first layer is the input layer, which is the features of the extracted current, vibration and acoustic emission signals; the second layer is the first hidden layer, with a total of 60 neurons, the dropout probability p is set to 0.5, the activation function is the tanh function, and the weights are initialized with normal distributed random numbers; the third layer is the second hidden layer, with a total of 20 neurons, the dropout probability p is 0.5, the activation function is the tanh function, and the weights are initialized with normal distributed random numbers; the fourth layer is the third hidden layer, with a total of 10 neurons, the activator is the tanh function, and the weights are initialized with normal distributed random numbers; the fifth layer is the output layer for the drill wear amount, and the activator is a linear function, such as Fig.10 Shown
[0156] Set the maximum number of training iterations to 1500, the batch size to 30, and use the Adam optimizer.
[0157] The results of the deep neural network were reduced through training. After multiple adjustments, the better model parameters were obtained: the number of neurons in the first hidden layer was 15, the dropout probability p was 0.5, and the number of neurons in the second hidden layer was 8, which further improved the monitoring accuracy of the model.
[0158] Since the extracted signal features were not screened before training the model, the contributions of the 263 signal features to the model were not the same. The model data was read by Keras, the weight value of the first layer was extracted, the absolute value of each feature was added and ranked, and the signal feature contribution ranking was obtained. In order to simplify the calculation amount and improve the operation speed of the model, the top 30 signal features were selected, the remaining features were eliminated, and retraining was performed.
[0159] The accuracy of the new model on the training set and test set is slightly reduced, but it is within an acceptable range. At the same time, the top 30 signal features do not include the features of the vibration sensor signal installed on the workpiece plate, indicating that it is not sensitive to drill wear and can be used without this sensor in industrial applications.
[0160] Theoretically, the wear of the drill bit blade can be calculated by real-time acquisition of the spindle current during the drilling process, thus realizing model-driven monitoring of drill bit wear. After the previous processing, the drilling signal of the drill bit is input into the model-driven monitoring model based on the current signal and the data-driven monitoring model based on the deep neural network. The monitoring values of the two are compared. Both models can reflect the change trend of drill bit wear, and the monitoring values fluctuate around the true value of the drill bit blade wear value VB after fitting. However, compared with the data-driven monitoring value, the model-driven monitoring value fluctuates more significantly, which is related to the fact that some ideal assumptions made when the model was established make the model inaccurate, and the collected current signal is easily disturbed by the environment. The data-driven monitoring model based on the deep neural network contains not only the information of the current signal, but also the information of the vibration signal and the acoustic emission signal, so the monitoring performance of the data-driven model is better.
[0161] Comparing the generalization capabilities of the model-driven monitoring model and the data-driven model, the present invention intends to integrate the model and data-driven drill wear monitoring method to improve the accuracy of drill wear monitoring and the versatility of the monitoring method, such as Fig.12 It is a schematic diagram of the drill wear monitoring model driven by model and data fusion.
[0162] 4. In step S4, an improved Kalman filtering algorithm is used to perform parallel fusion of the results of the model-driven model and the data-driven model to further improve the model accuracy.
[0163] The Kalman filter algorithm is a recursive prediction filter algorithm that has both filtering functions and the ability to predict data at the next moment. The algorithm is divided into two steps: estimation and correction. First, the current state is estimated based on the state at the previous moment, and then the optimal state is estimated by combining the current estimated state with the observed state.
[0164] In the estimation step, the expressions of the state estimation equation and the state estimation transfer equation are:
[0165] x k =Ax k-1 +Bu k +w k
[0166] z k =Hx k +v k
[0167] In the formula, x k is the system state vector, u k is the driving input vector, A is the state transfer matrix, B is the input control matrix, and w k is the estimated noise with variance W, z k is the real state x at time k kThe observation value, H is the observation matrix, v k is the observation noise with variance V. The prior estimation equation of the system state is:
[0168] x p,k =Ax k-1 +Bu k-1
[0169] In the formula, x p,k is the system state prediction value, which is the system state vector x at the previous moment k-1 and the driving input vector u k-1 Calculated.
[0170] The improved Kalman filtering algorithm applies the output of the data-driven model to the nonlinear prediction of the system state, and the calculation formula of the system state prediction value becomes:
[0171] x p,k =Ax f,k-1 +(x d,k -x d,k-1 )
[0172] In the formula, x f,k-1 is the output value of the fusion model at time k-1, x d,t -x d,t-1 Represents the change in predicted system state values at adjacent moments based on the data-driven model.
[0173] The model-driven drill wear monitoring model based on current signals has good generalization performance, and the model output value fluctuates around the actual value. The data-driven model based on deep neural networks has insufficient training sample mining, and the monitoring data will deviate in different test samples. The improved Kalman filtering algorithm converts the output of the model-driven model and the data-driven model into the observed value and the predicted value, respectively, and the observed value is the state value (observation matrix H = 1), and the state estimation equation and state estimation transfer equation become:
[0174] x k =x k-1 +(x d,k -x d,k-1 )+w k-1
[0175] z k =x m,k =x k +v k
[0176] In the formula, x m It is the output of the model-driven model, that is, the observed value. The solution of the Kalman filter algorithm can be calculated using recursive estimation, and the time update equation is:
[0177] x p,k =x h,k-1 +(x d,k -x d,k-1 )
[0178] R k - =R k-1 +W k-1
[0179] In the formula, x h is the optimal estimate of the system, and R is the covariance of the measurement noise. The state update equation is:
[0180] W k =R k - (R k - +V k-1 ) -1
[0181] x h,k =x p,k +K(x m,k -x p,k )
[0182] R k =(IK)R k -
[0183] Where K is the Kalman gain coefficient. Due to the unpredictability of the drill wear process, the noise characteristics of the system in the Kalman filter algorithm are time-varying. In order to improve the accuracy of the algorithm, when the algorithm is recursive, the variance W and V of the estimated noise and observation noise of the online estimation system are estimated by moving window technology (window length is n) as follows:
[0184]
[0185] According to the above formula, the fusion model parameters are corrected online with the goal of optimizing the model's estimation performance.
[0186] like Fig.13 The model based on improved Kalman filtering and data fusion driven modeling diagram, the system state (data driven model) predicted value x p and the model-driven model prediction value x m According to the prediction accuracy (W and V), the weights (1-K and K) are automatically adjusted online to obtain x h Finally, the fusion of model-driven model and data-driven model is realized. h This is the fusion output value.
[0187] The monitoring results of the model and the single model for the drill bit were then compared, and the root mean square error RMSE was used as the evaluation index of the model monitoring effect. For the drill bit data, the RMSE of the model data fusion model was slightly higher than that of the data-driven model, but for another set of drill bit data, the RMSE of the model data fusion model was lower than that of the data-driven model, indicating that the model data fusion model has the highest monitoring accuracy. The results show that compared with the single model-driven monitoring model or the data-driven monitoring model, the model and data fusion model comprehensively utilizes the advantages of the strong generalization of the model-driven model and the high monitoring accuracy of the data-driven model, improving the generalization ability of the model while ensuring the monitoring accuracy.
[0188] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the concept and technical solution of the present invention, or the practical invention and technical solution are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.
Claims
1. A method for monitoring wear of a carbon fiber reinforced composite drill bit, characterized in that: The following steps are involved: S1. Considering tool wear, a model-driven drill wear monitoring model based on current signal is established in combination with the electric spindle current sensor signal, which includes: derivation of drill main cutting edge load model, derivation of drill chisel edge load model, derivation of drilling model load considering drill wear, and modeling of drilling load-spindle current model; S2. Analyze the wear process of the drill bit to obtain the wear characteristics of the drill bit, and select multiple sensor signals as monitoring signals of the drill bit wear status in combination with the drilling environment and conditions, and perform feature extraction of the sensor signals in the time domain and frequency domain; S3. Using the extracted characteristic signals, a data-driven multi-signal fusion drill wear monitoring model based on deep neural network is established; S4. Combined with the model driven drill wear monitoring model based on current signal, a model based on improved Kalman filter algorithm and a data fusion driven drill wear monitoring model are established to realize the monitoring of drill wear during the drilling process.
2. A carbon fiber reinforced composite material drill bit wear monitoring method according to claim 1, characterized in that: In step S1, considering that it is difficult to establish an accurate analytical model for carbon fiber reinforced composite materials, a semi-empirical formula is used to model the drilling process: In the formula, F vu and F hu are the forces on the cutting edge unit in the vertical and horizontal directions, respectively; A and B are parameters related to drilling parameters and drill geometry to be determined experimentally; γ is the drill rake angle; and h is the cutting depth.
3. A carbon fiber reinforced composite material drill bit wear monitoring method according to claim 1, characterized in that: In the derivation of the load model of the main cutting edge of the drill bit in step S1, when the main cutting edge of the drill bit completely cuts into the workpiece material when drilling carbon fiber reinforced composite materials, it can be regarded as a series of adjusted unit orthogonal cutting. According to the mechanical analysis method and the microelement method, the main cutting edge is divided into a number of cutting units; At any cutting position angle θ, the cutting layer thickness h and the cross-sectional area of the cutting unit are: Where φ is half of the drill bit angle, f r The feed rate per revolution of the drill is: f r =60f / (1000n) (1-3) Where n is the spindle speed (krpm), f is the spindle feed speed (mm / s). For the convenience of calculation, the coordinate system is set to the polar coordinate system, and the radial distance is normalized: Where ρ is the radial coordinate in the polar coordinate system, θ(ρ) is the cutting edge inclination angle at the radial coordinate ρ, R is the drill radius, r is the distance from a point on the main cutting edge to the center, δr is the cutting edge unit width at the radial coordinate ρ, and the cosine value can be obtained from the drill geometry: Where b is the drill core thickness. Combined with the above formula, the force F generated by the entire cutting edge perpendicular to the cutting edge direction and the cutting speed direction is v And the cutting speed direction force F h It can be obtained by integrating the unit force: In the formula, the entire cutting edge is calculated, the upper limit of the integral is: j = R / R = 1, the lower limit of the integral is: i = (b / sinψ) / R, ψ is the chisel edge angle, and the axial force F T is the force F perpendicular to the two cutting edges v Axial force F a The resultant force, drilling torque M h is the integral of the moment acting on the cutting element, as follows: In the formula, the drill bit rake angle γ is divided into two parts: the drill bit rake angle itself - static rake angle γ s The second part is determined by the feed rate - dynamic rake angle γ d , as shown below: 。 4. A carbon fiber reinforced composite material drill bit wear monitoring method according to claim 1, characterized in that: In the derivation of the drill bit chisel edge load model in step S1, the micro-element force on the chisel edge can be expressed as: Where E3 is the Young's modulus of the workpiece material in the feed direction, v is the Poisson's ratio of the workpiece material, and γ ω is the half angle of the chisel edge, γ d It is the angle between the total cutting force of the chisel edge and the vertical direction, that is, the aforementioned dynamic rake angle. Due to the existence of the dynamic rake angle, the axial force on the chisel edge can be expressed as: 。 5. The method for monitoring wear of a carbon fiber reinforced composite material drill bit according to claim 3, characterized in that: In the load derivation of the drilling model considering drill wear in step S1, the cutting force caused by tool wear can be divided into two parts: F perpendicular to the tool wear area and vw and the friction force F parallel to the wear area hw , whose expression is as follows: Where H is the hardness of the material, VB is the width of the single-edge wear zone, s is the length of the wear zone, μ is the friction coefficient between the workpiece and the tool material, and the axial force F generated by drill wear is Tw F vw The axial force F aw Assuming that the two edges of the drill are worn evenly during drilling, the torque M generated by the wear of the drill flank is w The differential element method and double integral calculation are used in the wear area as follows: Where α is the back angle of the main cutting edge, l is the projection length of the main cutting edge on the xy plane, S w is the projection of the drill bit back face wear area on the xy plane. According to the above formula, the tool wear factor can be added to modify the formula of the axial force and torque of the drill bit, as shown in formula (1-13). According to this formula, the axial force and torque of the drill bit under different wear degrees can be calculated; 。 6. A carbon fiber reinforced composite material drill bit wear monitoring method according to claim 5, characterized in that: In the drilling load-spindle current modeling of step S1, the spindle current and the torque on the drill bit have the following relationship model: In the formula, K t is the torque constant of the spindle motor, I is the load current of the spindle motor, T m is the load torque acting on the spindle motor, T f is the internal friction torque overcome when the spindle motor rotates, J is the equivalent moment of inertia of the spindle, ω is the angular frequency of the spindle, M total is the torque on the drill bit, see formula (1-13); In the actual drilling process, under the control of the electric spindle driver, the rotation speed of the spindle motor remains unchanged, that is, the spindle angular acceleration is 0 (dω / dt=0), so the equivalent rotational inertia J of the spindle can be ignored. When the spindle motor is idling, it only needs to overcome the internal friction torque of the motor, and the torque on the drill bit is 0 (M total =0), then: K t I0(n)=T f (1-15) In the formula, I0(n) is the idling current when the spindle motor speed is n. Substituting formula (1-15) into formula (1-14), the drilling current increment △I is obtained as follows: ΔI=I-I0(n)=M total / K t (1-16) It can be seen from the formula that the spindle current increment is proportional to the load on the drill bit, thus establishing a theoretical model of drilling load and spindle current. Combined with the drill wear-drilling load model, the drill wear status can be monitored by collecting the real-time spindle current of the electric spindle during drilling. Substituting equations (1-7), (1-12), and (1-13) into (1-16), we obtain: From this formula, it can be seen that the drilling current increment △I of the high-speed spindle can be expressed by the drilling parameters (spindle speed n and feed speed f) and the single-edge wear band width VB, that is, △I=f(n,f,VB). A, B, μ, H, K in formula (1-17) and (1-13) are t , E3, v and other parameters are constants related to the drill geometry parameters, drilling parameters, workpiece material properties and electric spindle characteristics, and can be determined by fitting through variable parameter drilling load and spindle current measurement experiments.
7. A carbon fiber reinforced composite material drill bit wear monitoring method according to claim 1, characterized in that: In step S2, current signal feature extraction: the current signal of the drill bit when working is collected by a high-speed data acquisition card, and 15 time domain features are extracted; vibration signal feature extraction: two miniature three-axis acceleration sensors are placed on the spindle fixture and the workpiece, and the vibration signal is collected in combination with the NI signal acquisition system, and the time domain and frequency domain features are extracted, with a total of 229 features; acoustic emission signal feature extraction uses the SAEU2S acoustic emission system to collect acoustic emission signals with a sampling frequency of 1 MHz. The acoustic emission signals collected during the drilling process of drill bits with different degrees of wear are used to extract 19 time domain features, and a total of 263 features are extracted from the current signal, vibration signal and acoustic emission signal.
8. The method for monitoring wear of a carbon fiber reinforced composite material drill bit according to claim 1, characterized in that: In step S3, the extracted features are normalized to produce feature samples as input of the deep neural network, and the drill wear value is used as output for model training to establish a data-driven monitoring model based on the deep neural network.
9. A carbon fiber reinforced composite material drill bit wear monitoring method according to claim 1, characterized in that: In step S4, a deep neural network model (DNN) is used to realize data-driven multi-signal fusion monitoring of drill bit wear. Dropout and model parameter optimization are used to reduce the root mean square error of the model and improve the accuracy of the model. Finally, an improved Kalman filter algorithm is used to realize model and data fusion driven monitoring of drill bit wear.
Citation Information
Cited By
Abrasion detection method and system of drill bit based on stone drilling
CN120612331A
Tool durability detection and characterization method
CN120874021A
