Dynamic optimization method for surface quality of camshaft non-circular profile high speed grinding

Abstract

The present application relates to the technical field of machining process parameter optimization, and particularly relates to a surface quality dynamic optimization method for camshaft non-circular contour high-speed grinding; the method comprises the following steps: collecting multi-source mixed signals in the camshaft non-circular contour high-speed grinding process; separating out source signal components corresponding to the grinding process state by using a signal processing method based on WPD-EWT-FastICA; selecting a high-correlation feature set with high correlation with the grinding process state by using ReliefF; dividing state labels; training a recognition model based on CNN-LSTM-Attention; acquiring real-time multi-source mixed signals during processing, and outputting real-time grinding process states; triggering a dynamic response mechanism, updating a multi-objective optimization function, and acquiring an optimal process parameter combination; the method can adaptively adjust the optimal process parameter combination in real time, and dynamically optimize surface roughness, residual stress and material removal rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machining process parameter optimization, and particularly relates to a surface quality dynamic optimization method for high-speed grinding of a camshaft non-circular profile. BACKGROUND

[0002] A camshaft is a key component of an engine, and the machining quality of the working surface of the camshaft is a core factor affecting the service efficiency and service life of the engine. High-speed grinding, as the last process for the non-circular profile surface of the cam, determines the quality of the working surface.

[0003] However, the grinding process states such as grinding chatter, grinding wheel wear and workpiece surface burn are accompanied in the high-speed grinding process of the cam, and different grinding process states have different influences on the surface quality after grinding. If precise control of the surface quality after grinding is to be achieved, the change of the grinding process state needs to be considered when modeling the grinding surface quality and optimizing the process parameters. The change of the grinding process state will cause the optimal process parameters to be invalid, but the change of the grinding process state is uncertain and the dynamic change process thereof is not fully understood at present, which makes the optimization of the process parameters difficult.

[0004] In addition, when optimizing the process parameters, the machining efficiency (i.e. material removal rate) needs to be considered in addition to the surface quality (generally including surface roughness and residual stress) of the finished product, which makes it necessary to establish a multi-objective optimization model when optimizing the process parameters. The traditional process parameter optimization method is difficult to consider the dynamic change of the machining process and is difficult to meet the process parameter optimization demand of the complex and changeable machining environment. When the machining state changes, the static process parameter optimization method will not be suitable for the current machining.

[0005] In summary, there is no surface quality dynamic optimization method for high-speed grinding of a camshaft non-circular profile that can fully consider the dynamic multi-objective optimization after the change of the machining state in the prior art, and therefore it is necessary to propose a solution to the current situation. SUMMARY

[0006] The present application relates to the technical field of machining process parameter optimization, and particularly relates to a surface quality dynamic optimization method for high-speed grinding of a camshaft non-circular profile.

[0007] In order to achieve the above-mentioned purpose, the application provides a camshaft non-circular contour high-speed grinding surface quality dynamic optimization method, which comprises the following steps: S1. Collecting multi-source mixed signals in the camshaft non-circular contour high-speed grinding process through acoustic emission sensors and vibration sensors; obtaining intermediate signals after pre-processing the multi-source mixed signals, and separating the intermediate signals into source signal components corresponding to grinding process states by using a signal processing method based on WPD-EWT-FastICA, wherein the grinding process states include grinding chatter, grinding wheel wear and grinding burn; S2. Extracting time domain, frequency domain and time-frequency domain features of the source signal components separated in step S1, and selecting a high-correlation feature set with high correlation with the grinding process states by using ReliefF; S3. Quantitatively evaluating the states of grinding chatter, grinding wheel wear and grinding burn according to quantitative standards, and dividing different state labels; obtaining a high-correlation feature set for training by using steps S1 and S2, constructing a grinding process state recognition input matrix by using the high-correlation feature set for training and the state labels corresponding thereto, inputting the grinding process state recognition input matrix into a recognition model based on CNN-LSTM-Attention for training, and obtaining the trained recognition model based on CNN-LSTM-Attention; S4. Obtaining real-time multi-source mixed signals during actual processing, obtaining real-time high-correlation feature sets by processing the real-time multi-source mixed signals according to steps S1 and S2, inputting the real-time high-correlation feature sets into the trained recognition model based on CNN-LSTM-Attention, and outputting real-time grinding process states; S5. Triggering a dynamic response mechanism according to the real-time grinding process states output in step S4, stopping processing if it belongs to a serious influence state, and selecting corresponding historical experimental and simulation data according to the real-time grinding process state if it is a non-serious state that can be optimized, and constructing a multi-objective optimization function with surface roughness R a , residual stress σ and material removal rate MRR as optimization objectives, and grinding wheel linear speed v s , workpiece speed v w and grinding depth a p as decision variables, solving the multi-objective optimization function by using an NSGA-III algorithm, obtaining a Pareto optimal solution set, and determining an optimal process parameter combination according to production requirements; S6. When the real-time grinding process state changes, the multi-objective optimization function is dynamically updated, and the optimal process parameter combination is reacquired.

[0008] Further, the preprocessing of the multi-source mixed signal in step S1 includes an outlier detection and optimization process: first, the original signal is divided into multiple signal segments; then, the signal segments are further divided into signal parts, and if the root mean square value of a signal part exceeds a set threshold, the signal part is regarded as an outlier, and the threshold is obtained according to the 3σ principle; finally, the outliers are deleted from the original signal and the data is supplemented by linear interpolation.

[0009] Further, the signal processing method based on WPD-EWT-FastICA in step S1 includes: processing the acoustic emission signal by using the WPD-FastICA method, the sampling frequency is 1 MHz, the wavelet packet decomposition is performed for 3 layers, the wavelet basis function is selected as db10 wavelet, the node signal with the highest energy ratio is selected from the decomposed node signal to construct an input matrix, and the independent component is separated by the FastICA algorithm; the vibration signal is processed by using the EWT-FastICA method, the sampling frequency is 6400 Hz, the signal is decomposed by using the empirical wavelet transform, the Fourier spectrum of the signal is adaptively segmented, the frequency band boundary is determined according to the maximum value point of the spectrum, the empirical wavelet filter bank is constructed, the input matrix is constructed by the decomposed components, and the independent component is separated by the FastICA algorithm; the independent component with a frequency peak near 300 kHz is taken as the grinding burn source signal component, the independent component with a frequency near 870 Hz is taken as the grinding chatter source signal component, and the independent component with a frequency peak near 450 kHz is taken as the grinding wheel wear source signal component.

[0010] Further, in step S3: the grinding chatter state label includes stable grinding state, slight chatter state and severe chatter state, the grinding wheel wear state includes initial wear state, normal wear state and severe wear state, and the grinding burn state includes no burn state, slight burn state and severe burn state.

[0011] Further, in step S3: the grinding chatter state label is divided by: when the vibration signal spectrum energy is uniformly distributed and there is no vibration mark on the workpiece surface, it is the stable grinding state; when the vibration signal spectrum energy starts to gather at 870 Hz and irregular grinding marks appear on the workpiece surface, it is the slight chatter state; when the vibration signal spectrum energy is concentrated at 870 Hz and vibration marks are formed on the workpiece surface, it is the severe chatter state; the grinding wheel wear state label is divided by: detecting the abrasive wear area ratio R area , when 0 < R area < 16%, it is the initial wear state, when 16% ≤ R area < 29%, it is the normal wear state; and when R area≥ 29% is a severe wear state; the classification method of the grinding burn state label is to detect the surface Vickers hardness and the degree of discoloration, when HV ≥ 550 and the surface is bright yellow or no discoloration, it is a non-burn state; when 470 < HV < 550 and the surface is yellow-black or slightly discolored, it is a slight burn state; when HV ≤ 470 and the surface is blue-black with cracks, it is a severe burn state.

[0012] Further, the dynamic response mechanism in step S3 is: if the output real-time grinding process state is a slight burn state, a severe burn state, a severe wear state or a severe chatter state, the machining is stopped; if the output real-time grinding process state is a normal wear state or a slight chatter state, the multi-objective optimization function is dynamically updated according to the current state combined with historical experimental and simulation data, the NSGA-III algorithm is used to solve the updated multi-objective optimization function, and the optimal process parameter combination suitable for the current state is output; when the normal wear state and the slight chatter state exist at the same time, the dynamic updating mechanism of the slight chatter state is given priority.

[0013] Further, the recognition model based on CNN-LSTM-Attention sequentially includes a convolutional neural network module, an attention mechanism module and a long short-term memory network module; the convolutional neural network module is used to receive the input sensor time series signal, the time series signal is composed of data points at consecutive time steps; a one-dimensional convolution operation is performed to slide a convolution kernel in the time dimension, the convolution kernel covers multiple time steps to form a local time window, spatial features in each local time window are extracted, and a feature sequence composed of time step features is output; the attention mechanism module is used to process the feature sequence, calculate the attention weight of each time step, and multiply the attention weight with the original feature of the corresponding time step element by element to enhance the feature representation of the key time step, and generate a weighted feature sequence; the long short-term memory network module is used to receive the weighted feature sequence, capture the long-term dependence between time steps, and output a final feature vector; the final feature vector is input into a Softmax classifier to obtain the grinding process state.

[0014] Further, the multi-objective optimization function in step S5 includes:

[0015] ,

[0016] ,

[0017] ,

[0018] In the above three formulas, R a is the roughness; σ is the residual stress; Z wMRR is the material removal rate of single abrasive grain, which is used to calculate the total material removal rate MRR;v w V is the workpiece linear speed;v s V is the grinding wheel linear speed;a p H is the grinding depth; K x, y, z, a, b, c, d, e, f are obtained by fitting experimental data; τ is the ratio of the true contact arc length to the motion contact arc length; k is the ratio of the actual grinding depth to the theoretical grinding depth; θ is the abrasive grain top angle;d s R is the radius of the grinding wheel;d w R is the radius of the workpiece; for (v s ±v w ), select + when grinding in reverse, and select - when grinding in forward.

[0019] Further, the constraint condition in step S5 is:

[0020] .

[0021] Compared with the prior art, the camshaft non-circular contour high-speed grinding surface quality dynamic optimization method provided by the application realizes high-precision separation of multi-source sensing signals by adopting a signal processing method based on WPD-EWT-FastICA, accurately extracts source signal components corresponding to grinding chatter, grinding wheel wear and grinding burn, significantly reduces noise interference, improves the signal-to-noise ratio of state recognition, and improves the monitoring accuracy; then, through the recognition model based on CNN-LSTM-Attention, combined with the dynamic response mechanism and the NSGA-III multi-objective optimization algorithm, the method can adaptively adjust the optimal process parameter combination according to the real-time grinding process state, dynamically optimize the surface roughness, residual stress and material removal rate, and solve the problem that the static process parameter optimization method in the prior art cannot cope with the change of the grinding process state, resulting in invalidation of the optimal process parameter combination, which provides a new technical idea and research foundation for processing quality control. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a flowchart of the camshaft non-circular contour high-speed grinding surface quality dynamic optimization method;

[0023] Figure 2 is a flowchart of the signal processing method based on WPD-EWT-FastICA;

[0024] Figure 3 is the network structure of CNN-LSTM-Attention;

[0025] Figure 4 is a flowchart of NSGA-III;

[0026] Figure 5 is a processing flow chart of the recognition model based on CNN-LSTM-Attention;

[0027] Figure 6 is the accuracy rate statistics of the grinding wheel wear recognition result based on CNN-LSTM-Attention;

[0028] Figure 7 is the accuracy rate statistics of the grinding wheel wear recognition result based on CNN;

[0029] Figure 8 is the accuracy rate statistics of the grinding wheel wear recognition result based on LSTM;

[0030] Figure 9 is the accuracy rate statistics of the grinding wheel wear recognition result based on CNN-LSTM. DETAILED DESCRIPTION

[0031] The embodiments of the application are described in detail below.

[0032] The embodiment provides a surface quality dynamic optimization method for camshaft non-circular contour high-speed grinding, like Figure 1As shown, it includes the following steps: S1. Collecting multi-source mixed signals in the camshaft non-circular contour high-speed grinding process through acoustic emission sensors and vibration sensors; obtaining intermediate signals after preprocessing the multi-source mixed signals, and separating the intermediate signals into source signal components corresponding to the grinding process states by using a signal processing method based on WPD (Wavelet Packet Decomposition, Wavelet Packet Decomposition)-EWT (Empirical Wavelet Transform, Empirical Wavelet Transform)-FastICA (Fast Independent Component Analysis, Fast Independent Component Analysis), wherein the grinding process states include grinding chatter, grinding wheel wear and grinding burn; S2. Extracting time domain, frequency domain and time-frequency domain features from the source signal components separated in step S1, and selecting a high-correlation feature set with high correlation with the grinding process states by ReliefF; S3. Quantitatively evaluating the states of grinding chatter, grinding wheel wear and grinding burn according to the quantification standard, and dividing different state labels; obtaining a high-correlation feature set for training through steps S1 and S2, constructing a grinding process state recognition input matrix by using the high-correlation feature set for training and the state labels corresponding thereto, and inputting the grinding process state recognition input matrix into a recognition model based on CNN (Convolutional Neural Networks, Convolutional Neural Networks)-LSTM (Long Short-Term Memor, Long Short-Term Memory Network)-Attention (Attention Mechanism) for training to obtain a trained CNN-LSTM-Attention recognition model; S4. Obtaining real-time multi-source mixed signals during actual processing, obtaining real-time high-correlation feature sets after processing the real-time multi-source mixed signals through steps S1 and S2, inputting the real-time high-correlation feature sets into the trained CNN-LSTM-Attention recognition model, and outputting real-time grinding process states; S5. Triggering a dynamic response mechanism according to the real-time grinding process states output in step S4, stopping processing if it belongs to a serious influence state, and selecting corresponding historical experimental and simulation data according to the real-time grinding process states if it is a non-serious state that can be optimized, and constructing a surface roughness R a , residual stress σ and material removal rate MRR as optimization objectives, and the grinding wheel linear speed v s , workpiece speed v w and grinding depth a pFor the multi-objective optimization function of the decision variable, the NSGA-III algorithm (Non-Dominated Sorting Genetic Algorithm III) is used to solve the multi-objective optimization function, obtain the Pareto optimal solution set, and determine the optimal process parameter combination according to the production requirements; S6, when the real-time grinding process state changes, dynamically update the multi-objective optimization function, and reacquire the optimal process parameter combination.

[0033] Based on the above step setting, the surface quality dynamic optimization method for camshaft non-circular contour high-speed grinding adopts the WPD-EWT-FastICA-based signal processing method to realize high-precision separation of multi-source sensing signals, accurately extract source signal components corresponding to grinding chatter, grinding wheel wear and grinding burn, which significantly reduces noise interference, improves the signal-to-noise ratio of state recognition, and improves the monitoring accuracy. Then, through the recognition model based on CNN-LSTM-Attention, the grinding burn, grinding chatter and grinding wheel wear state changes are identified online, and the dynamic response mechanism and NSGA-III multi-objective optimization algorithm are combined, so that the method can adaptively adjust the optimal process parameter combination according to the real-time grinding process state, dynamically optimize the surface roughness, residual stress and material removal rate, and solve the problem that the static process parameter optimization method in the prior art cannot cope with the change of the grinding process state, resulting in the failure of the optimal process parameter combination. It provides a new technical idea and research foundation for processing quality control.

[0034] Next, each step in the method will be further described in detail.

[0035] In this embodiment, the acoustic emission data acquisition platform uses a W800 acoustic emission sensor and a PAS preamplifier, with a frequency bandwidth of 50-800 kHz. The preamplification ratio is set to 40 dB, and the filter frequency band is 20-1200 kHz. The acquisition card is USB-6351 from the United States NI company. The attenuation of acoustic emission signals is mainly affected by the propagation distance, so the installation of acoustic emission sensors needs to be as close to the grinding area as possible without interfering with the grinding process. In this embodiment, the acoustic emission sensor is installed in the Y-axis direction of the grinder tailstock, and the acoustic emission signal sampling frequency is set to 1 MHz. The vibration signal data acquisition platform uses the Yiheng ECON vibration data acquisition system. In order to obtain accurate grinding process vibration state information, the vibration sensor is installed in the headstock X direction, and the grinding vibration frequency is mainly concentrated in 0-3000 Hz, so the acquisition frequency is set to 6400 Hz.

[0036] In the embodiment, the preprocessing of the multi-source mixed signal in step S1 includes signal chopping, truncation, time registration and the like to enhance the readability of the data. To avoid aliasing of the reconstructed signal, the camshaft high-speed grinding processing data acquisition is based on the Nyquist-Shannon sampling theorem, i.e. s ≥ 2f max , combined with experimental research, the acoustic emission signal sampling frequency is set to 1 MHz, and the vibration signal sampling frequency is set to 6400 Hz. In order to obtain effective sensor data, the incomplete time sequence signal waveform is removed by chopping. Due to the high sampling frequency and long acquisition time of the signal acquisition during the camshaft grinding process, the amount of grinding processing data is particularly large and difficult to analyze and process. Considering the curvature change and the sampling frequency, the camshaft lift and return are the weak parts of the surface quality, and grinding burn and grinding chatter usually occur first at this point, so the acoustic emission signal is truncated for 0.1 seconds and the vibration signal is truncated for 0.5 seconds for analysis. In addition, the base circle and the peach tip are also the positions that need to be paid attention to during camshaft grinding, so the same length of signal is also truncated for analysis. The multi-sensor information fusion needs to ensure the start time and end time of the signal truncated by multiple sensors. In the actual grinding experiment, multiple sensors may collect signals at the same time, which may cause inconsistency between signals of each sensor. Therefore, the time node needs to be set to correct the multi-sensor information to ensure that the analyzed data is consistent in time sequence.

[0037] In the complex environment of camshaft non-circular contour high-speed grinding, some abnormal signals may be collected, which will have a bad influence on the state monitoring of the grinding process. Therefore, a statistical-based method is constructed to detect and delete the abnormal values of the data and retain the normal signals. In the embodiment, the preprocessing of the multi-source mixed signal in step S1 includes outlier detection and optimization process: first, the original signal is divided into multiple signal segments; then, the signal segment is further divided into several signal parts, if the root mean square value of a signal part exceeds the set threshold value, the signal part is regarded as an outlier, and the threshold value is obtained according to the 3σ principle; finally, the outlier is deleted from the original signal and the data is supplemented by linear interpolation.

[0038] According to the principle, the threshold value is obtained by the following formula: ; wherein T p represents the threshold value of the root mean square of the signal p part; μ p and σ pThe average value and standard deviation of all signals before p in a signal part are represented respectively. The outliers of the vibration and acoustic emission data are removed and the data is supplemented by linear interpolation to ensure the continuity of the data in the time dimension. At the same time, the signal is denoised by wavelet thresholding, which further improves the signal-to-noise ratio of the signal, filters out the interference signal and makes the signal carrying useful information clearer.

[0039] The vibration signal and acoustic emission signal collected by the sensor in the grinding process are actually the joint action of various "sources" (grinding wheel wear, grinding chatter, grinding burn, machining noise and environmental noise, etc.) on the whole system. Therefore, the signals collected in the grinding process contain not only the information reflecting the state change of the grinding process, but also a large amount of noise. However, various states coexist in the actual grinding process, and it is difficult to predict the source signal and the mixing mode. The blind source separation method can separate the target component needed from the sensor data, but it must satisfy that the number of sensors is not less than the number of signal sources. However, it is difficult to install a large number of sensors in the grinding system, so the signal decomposition method can be combined to separate the sensor signal into multiple components, and then the blind source separation technology is used to separate each source signal. Therefore, considering the state of high-speed grinding of the camshaft non-circular contour in this embodiment, a signal separation method based on WPD-EWT-FastICA is proposed.

[0040] It is found through research that there is no obvious connection between grinding burn and vibration signal in time domain, frequency domain and time-frequency domain, and it is difficult to characterize the grinding burn state. This is because the sampling frequency of the vibration signal is low, and high-frequency elastic waves are usually generated during grinding burn, so it is difficult to monitor the information related to grinding burn. The sampling frequency of acoustic emission is high, which can better monitor the grinding burn. The vibration signal more reflects the contact between the grinding wheel and the workpiece, the vibration of the processing system and the environmental noise, etc., and the vibration signal is the most direct signal reflecting the vibration of the processing system / grinding chatter. Therefore, in this embodiment, the acoustic emission signal data is used for the analysis of grinding burn, and the vibration signal is used for the analysis related to grinding chatter / vibration. Further research shows that the signal containing only grinding wheel wear information is mainly concentrated in the frequency domain of 500-800 kHz. By analyzing the characteristics of grinding chatter, the signal containing only grinding chatter information is separated, and the frequency of grinding chatter is mainly concentrated in the frequency band of 870 Hz. By laser-induced separation, the signal containing only grinding burn information is separated, and the frequency peak of grinding chatter is near 300 kHz.

[0041] In this embodiment, as shown in Figure 2As shown, the signal processing method based on WPD-EWT-FastICA in step S1 includes: processing the acoustic emission signal by using the WPD-FastICA method, the sampling frequency is 1MHz, the wavelet packet decomposition is performed for 3 layers, the wavelet base function is selected as db10 wavelet, the node signal with the highest energy ratio is selected from the decomposed node signal to construct an input matrix, and the independent component is separated by the FastICA algorithm; the vibration signal is processed by using the EWT-FastICA method, the sampling frequency is 6400Hz, the empirical wavelet transform is used for decomposition, the Fourier spectrum of the signal is adaptively segmented, the frequency band boundary is determined according to the maximum value point of the spectrum, the empirical wavelet filter bank is constructed, the input matrix is constructed by using the decomposed component, and the independent component is separated by the FastICA algorithm; the independent component with the frequency peak value near 300kHz is taken as the grinding burn source signal component, the independent component with the frequency near 870Hz is taken as the grinding chatter source signal component, and the independent component with the frequency peak value near 450kHz is taken as the grinding wheel wear source signal component.

[0042] Based on the above step 1, the signal processing method based on WPD-EWT-FastICA can effectively separate the source signals related to the grinding process state, provide convenience for subsequent grinding process state recognition, and improve the accuracy of the recognition result.

[0043] In the embodiment, in step S3: the grinding chatter state label includes a stable grinding state, a slight chatter state and a severe chatter state, the grinding wheel wear state includes an initial wear state, a normal wear state and a severe wear state, and the grinding burn state includes a non-burn state, a slight burn state and a severe burn state. Further, the embodiment also provides a division method of various state labels. The division method of the grinding chatter state label is that when the vibration signal spectrum energy is uniformly distributed and the workpiece surface is free of vibration marks, it is a stable grinding state; when the vibration signal spectrum energy starts to gather at 870Hz and the workpiece surface appears irregular grinding marks, it is a slight chatter state; and when the vibration signal spectrum energy is concentrated at 870Hz and the workpiece surface forms vibration marks, it is a severe chatter state; the division method of the grinding wheel wear state label is that when the ratio R area , 0 < R area < 16% is an initial wear state, 16% ≤ R area < 29% is a normal wear state; and R area≥ 29% is a severe wear state; the classification method of the grinding burn state label is to detect the surface Vickers hardness and the degree of discoloration, when HV ≥ 550 and the surface is bright yellow or no discoloration, it is a non-burn state; when 470 < HV < 550 and the surface is yellow-black or slightly discolored, it is a slight burn state; when HV ≤ 470 and the surface is blue-black with cracks, it is a severe burn state.

[0044] In this embodiment, the dynamic response mechanism in step S3 is: if the output real-time grinding process state is a slight burn state, a severe burn state, a severe wear state or a severe chatter state, the machining is stopped; if the output real-time grinding process state is a normal wear state or a slight chatter state, the multi-objective optimization function is dynamically updated according to the current state combined with historical experimental and simulation data, the NSGA-III algorithm is used to solve the updated multi-objective optimization function, and the optimal process parameter combination suitable for the current state is output; when the normal wear state and the slight chatter state exist at the same time, the dynamic updating mechanism of the slight chatter state is given priority, because the transition from the slight chatter state (chatter incubation period) to the severe chatter state is particularly serious.

[0045] In order to avoid the limitation of single model performance, multi-model fusion has become a means to improve model performance. In order to meet the accurate identification of the camshaft high-speed grinding process state, a grinding process state identification model is constructed by using a fusion model, so as to improve the identification accuracy and generalization ability of the model. Through the combination of CNN and LSTM, the problems of low identification accuracy and insufficient generalization ability of the machining state under complex working conditions can be effectively solved. Through further feature extraction by CNN, the influence of working conditions on signal features is reduced, and the relevance between the selected features and the grinding process state is further excavated. LSTM can maintain short-term memory, suppress irrelevant information and avoid the problem of gradient disappearance. However, CNN-LSTM cannot distinguish the importance of sequences. Therefore, Attention is introduced to select the information to be focused on from the data. Attention only focuses on the part closely related to the task, enhances the feature recognition ability, and has strong compatibility, which significantly improves the performance and generalization ability of the model. It can be seen that the method of combining CNN and LSTM and adding Attention can greatly improve the performance of the network model.

[0046] In this embodiment, the features with low or no correlation with the grinding process state are removed by ReliefF to improve the generalization ability of the model; Relief calculates the corresponding weight according to the correlation between the features and the labels, but cannot handle multi-classification problems, while ReliefF can decompose multi-classification into multiple binary classification and use k-nearest neighbor method to solve the influence of noise. Specifically, 22 signal features of the grinding wheel wear signal are separated based on WPD-EWT-FastICA, and then a feature subset T wheel is constructed. Based on ReliefF, the features of the vibration signal and the acoustic emission signal are optimized, and the features with weight score greater than 0.01 are selected as the input feature matrix R wheel of the grinding wheel wear state recognition model. The weight score is calculated as shown in Table 1.

[0047] Table 1 ReliefF weight score of grinding wheel wear signal features

[0048]

[0049] 11 signal features of the grinding burn signal are separated based on WPD-EWT-FastICA, and then a feature subset T burn is constructed. Based on ReliefF, the features of the vibration signal and the acoustic emission signal are optimized, and the features with weight score greater than 0.01 are selected as the input feature matrix R burn of the grinding wheel wear state recognition model. The weight score is calculated as shown in Table 2.

[0050] Table 2 ReliefF weight score of grinding burn signal features

[0051]

[0052] 11 signal features of the grinding burn signal are separated based on WPD-EWT-FastICA, and then a feature subset T chatter is constructed. Based on ReliefF, the features of the vibration signal and the acoustic emission signal are optimized, and the features with weight score greater than 0.01 are selected as the input feature matrix R chatter of the grinding wheel wear state recognition model. The weight score is calculated as shown in Table 3.

[0053] Table 3 ReliefF weight score of grinding burn signal features

[0054]

[0055] In the embodiment, the recognition model based on CNN-LSTM-Attention sequentially comprises a convolutional neural network module, an attention mechanism module and a long short-term memory network module; the convolutional neural network module is used for receiving an input sensor time series signal, the time series signal being composed of data points at continuous time steps; a one-dimensional convolution operation is performed to slide a convolution kernel in the time dimension, the convolution kernel covering multiple time steps to form a local time window, spatial features within each local time window are extracted, and a feature sequence composed of time step features is output; the attention mechanism module is used for processing the feature sequence, calculating an attention weight of each time step, and multiplying the attention weight with the original feature of the corresponding time step element by element to enhance the feature representation of a key time step and generate a weighted feature sequence; the long short-term memory network module is used for receiving the weighted feature sequence, capturing long-term dependencies between time steps, and outputting a final feature vector; the final feature vector is input into a Softmax classifier to obtain the grinding process state, and the network structure is as shown in Figure 3 The recognition model realizes the balance of efficiency, robustness and interpretability in the time series signal classification task through the CNN-LSTM-Attention hybrid architecture.

[0056] The process parameters of camshaft non-circular contour high-speed grinding have a great influence on the surface quality and present strong coupling and nonlinearity. For example, a moderate v s , low v w , lower a p , the better grinding surface quality can be obtained, but the material removal rate is reduced. At present, the selection of grinding process parameters is mostly realized by experience, and it is difficult to guarantee the grinding surface quality and efficiency of the camshaft. Therefore, in the embodiment, the grinding surface roughness R a and the surface residual stress σ are taken as the quality optimization objectives, the material removal rate MRR is considered, the grinding wheel linear speed v s , the workpiece speed v w and the grinding depth a p are taken as decision variables, the NSGA-III algorithm is used to obtain the multi-objective optimization process parameter combination of camshaft non-circular contour high-speed grinding, and the multi-objective optimization process parameter combination is used for the control of the part surface quality. In the grinding process, it is hoped to find an optimal balance point among various optimization objectives, and the solution of the multi-objective optimization problem is an optimal solution set, which is a Pareto optimal solution set or a non-inferior solution set method. The NSGA-III has obvious convergence promotion under specific multi-objective problems, adopts adaptive scaling mutation and increasing crossover factor strategy and reference point method based on dynamic crowding degree operator for selection operation, is suitable for multi-objective process parameter optimization of camshaft high-speed grinding surface quality, and the optimization process is as shown in Figure 4 .

[0057] In actual production, it is required to ensure the surface quality of the ground material while maintaining high grinding efficiency. To achieve efficient and high-quality grinding of the non-circular profile of the camshaft, multi-objective optimization is necessary, including constructing the objective function, selecting decision variables, and defining constraints.

[0058] In this embodiment, the multi-objective optimization function in step S5 includes:

[0059] (1) Objective function 1-roughness R a The empirical formula for surface roughness in grinding is expressed as a power function of the grinding process parameters. The empirical formula for surface roughness in grinding is:

[0060] ;

[0061] (2) Objective function 2 - residual stress σ. Experimental and simulation analysis show that the absolute value of residual stress in the X direction is usually greater than that in the Y direction, and both are compressive stresses. Therefore, to improve the efficiency and complexity of process parameter optimization, a functional relationship between the X-direction surface residual compressive stress σ and process parameters is constructed based on a multiple linear regression model:

[0062] ;

[0063] (3) Objective function 3 - Material removal rate (MRR), where material removal rate is the volume of material removed per unit time, representing grinding efficiency. Therefore, a calculation model for the material removal rate of a single abrasive grain is established by combining the shape of the abrasive grain and the material removal characteristics:

[0064] ;

[0065] In the above three formulas, R a Roughness; σ is residual stress; Z w v represents the material removal rate per abrasive grain, used to calculate the total material removal rate (MRR). w v is the linear velocity of the workpiece. s a is the linear velocity of the grinding wheel; p This refers to the grinding depth. K x, y, z, a, b, c, d, e, f are obtained by fitting experimental data; τ is the ratio of the actual contact arc length to the moving contact arc length; k is the ratio of the actual grinding depth to the theoretical grinding depth; θ is the abrasive grain tip angle; d s d is the radius of the grinding wheel; w Let (v) be the radius of the workpiece; for (v) in the formula s ±v w When grinding against the grain, select +; when grinding with the grain, select -.

[0066] In multi-objective optimization problems, constraints serve two main purposes: first, to improve the convergence speed of the optimization algorithm, reduce the search space, increase the efficiency of the algorithm, and reduce the waste of computational resources; second, to control the range of values ​​for decision variables to ensure that the optimization results meet actual requirements. The ultimate goal of a multi-objective optimization model is to obtain a set of Pareto optimal solutions, and all optimal solutions in the optimal solution set need to conform to the actual situation. Therefore, it is necessary to impose constraints on the decision variables. This paper imposes constraints on the decision variables based on a comprehensive consideration of the actual machining requirements of the camshaft and the performance of the machine tool. Therefore, in this embodiment, the constraint in step S5 is as follows:

[0067] .

[0068] Verification and Analysis of Online Monitoring of Grinding Process Conditions

[0069] The flowchart of the online monitoring method for grinding process status based on CNN-LSTM-Attention is shown in Figure 5. The input feature matrix R constructed as described in this embodiment... wheel R burn R chatter The data were used as input matrices to the CNN-LSTM-Attention-based recognition model, resulting in 724 sets of data. These sets were then used as training and testing sets for experimental validation. The specific model parameter settings are shown in Table 4 below.

[0070] Table 4 Model Parameter Settings

[0071]

[0072] like Figure 6 As shown, the accuracy rate for identifying grinding wheel wear on the training set is 99.5%, and the accuracy rate on the test set is 95.5%, indicating that the model has a high accuracy rate in identifying grinding wheel wear. Furthermore, for the identification of grinding burns and grinding chatter, the accuracy rates on both the training and test sets are greater than 94%, demonstrating that the model proposed in this embodiment has high identification accuracy.

[0073] The CNN-LSTM-Attention-based recognition model in this embodiment is compared and analyzed with CNN, LSTM, and CNN-LSTM. Here, only a dataset constructed using grinding wheel wear signal features is analyzed to demonstrate the model's superiority; the comparative experiments on grinding chatter and grinding burn monitoring models are not repeated. The parameter settings for the CNN, LSTM, and CNN-LSTM recognition models are shown in Table 5 below, and they use the same dataset as the CNN-LSTM-Attention-based recognition model in this embodiment.

[0074] Table 5 Parameter settings for the comparison models

[0075]

[0076] from Figure 7 It can be seen that the CNN training set has already exhibited overfitting, and its accuracy on the test set is lower than in this embodiment. From Figure 8 It can be seen that the accuracy of LSTM on both the test and training sets is lower than that of this embodiment. Since accurate identification of the grinding process state is a crucial prerequisite for dynamic process parameters, higher demands are placed on the accuracy and efficiency of the identification model. As shown in Figure 9, when using the CNN-LSTM identification model to predict grinding wheel wear, it was found that the training and prediction time of this model is longer than other models. For the same 1000 iterations, it takes 34 seconds, while other models take about 10 seconds. Furthermore, the identification results show that the accuracy of the CNN-LSTM identification model on the training set is lower than that on the test set, indicating a clear overfitting phenomenon. After comparing the four deep learning models, overall, the CNN-LSTM-Attention-based identification model in this embodiment demonstrates the advantages of a fusion model and can effectively identify changes in the grinding process state.

[0077] Results and Analysis of Grinding Surface Quality Optimization

[0078] Since a dynamic multi-objective optimization process consists of a series of static multi-objective optimization processes, the feasibility of the entire dynamic multi-objective optimization process can be proven by proving the feasibility of one of the static multi-objective optimization processes.

[0079] To this end, actual experiments were conducted for this embodiment, totaling 16 groups, and the average value of the target results under the 16 groups of experiments was obtained, i.e., the 16 experimental results. Then, based on experience, the optimal experimental parameters were selected from the 16 groups of experiments to obtain the target results under the optimal experimental parameters, i.e., the optimal experimental results. Finally, the optimal process parameters were obtained according to the method provided in this embodiment, and the target results under the optimal process parameters were obtained, i.e., the optimal optimization results. The obtained target results are shown in Table 6.

[0080] Table 6 Comparison of Target Results

[0081]

[0082] By comparing the optimal optimization results with the results of 16 sets of experiments, it can be seen that the surface roughness was optimized from 0.175 μm to 0.115 μm, an improvement of approximately 34.3%, which is a significant quality improvement, meaning a substantial increase in the surface finish of the workpiece. The residual stress was optimized from -485.39 MPa to -352.34 MPa, a reduction of approximately 27.4% in absolute value. This indicates a significant reduction in residual compressive stress on the workpiece surface, which is beneficial for improving the fatigue life and dimensional stability of the parts. The material removal rate decreased from 0.058 mm³ / s to 0.039 mm³ / s, sacrificing some grinding efficiency, but this was compensated by the significant improvement in the aforementioned quality indicators, which is well worth the effort for grinding the non-circular contour of camshafts with high quality requirements.

[0083] By comparing the optimal optimization results with the optimal experimental results, it can be seen that the surface roughness was optimized from 0.156 μm to 0.115 μm, an improvement of approximately 26.3%. The residual stress was optimized from -423.88 MPa to -352.34 MPa, a reduction of approximately 16.9% in absolute value. The material removal rate increased from 0.022 mm³ / s to 0.039 mm³ / s, and the grinding efficiency improved by 77.3%. It is evident that the optimization in this embodiment significantly improves grinding efficiency while also enhancing machining quality. This demonstrates that this embodiment can explore parameter combinations that are difficult to detect manually, achieving a multi-objective optimization effect.

[0084] Finally, the optimal process parameters were used in an actual machining experiment of non-circular profile grinding of the camshaft. The surface roughness, residual stress and material removal rate during the actual machining experiment were measured and calculated to obtain the experimental results. Then, the relative error of the optimal optimization results was calculated, as shown in Table 7 below. The overall relative error was controlled at around 10%, which verified the accuracy of this embodiment.

[0085] Table 7 Comparison of Target Results

[0086]

[0087] In summary, this method can adaptively adjust the optimal combination of process parameters according to the real-time grinding process status, and dynamically optimize surface roughness, residual stress and material removal rate. It solves the problem that the static process parameter optimization method in the prior art cannot cope with the changes in the grinding process status, which leads to the failure of the optimal combination of process parameters.

[0088] Where there is no conflict, the above embodiments and features can be combined with each other.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the preferred technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the present invention.

Claims

1. A method for dynamically optimizing the surface quality of a camshaft with a non-circular profile during high-speed grinding, characterized in that, Includes the following steps: S1. Acquire multi-source mixed signals during the high-speed grinding process of the non-circular profile of the camshaft using acoustic emission sensors and vibration sensors; After preprocessing the multi-source mixed signal, an intermediate signal is obtained. The signal processing method based on WPD-EWT-FastICA is used to separate the intermediate signal into source signal components corresponding to the state of the grinding process, which includes grinding chatter, grinding wheel wear and grinding burn. S2. Extract time-domain, frequency-domain, and time-frequency-domain features from the source signal components separated in step S1, and use ReliefF to select a set of highly correlated features that are highly correlated with the state of the grinding process; S3. Quantify and evaluate the various states of grinding chatter, grinding wheel wear, and grinding burn according to the quantification standard, and divide them into different state labels; obtain the high correlation feature set for training through steps S1 and S2, construct the high correlation feature set for training and its corresponding state labels into a grinding process state recognition input matrix, and input it into the recognition model based on CNN-LSTM-Attention for training to obtain the trained recognition model based on CNN-LSTM-Attention; S4. During actual processing, acquire real-time multi-source mixed signals, process the real-time multi-source mixed signals according to steps S1 and S2 to obtain real-time high-correlation feature sets, input the real-time high-correlation feature sets into the trained CNN-LSTM-Attention-based recognition model, and output the real-time grinding process status. S5. Based on the real-time grinding process status output in step S4, trigger the dynamic response mechanism. If it is a severely affected state, stop processing; if it is a non-severe state that can be optimized, select the corresponding historical experimental and simulation data based on the real-time grinding process status to construct a system with surface roughness R... a The residual stress σ and the material removal rate MRR are the optimization objectives, with the grinding wheel linear speed v as the optimization target. s Workpiece rotation speed v w and grinding depth a p The multi-objective optimization function is used as the decision variable. The NSGA-III algorithm is used to solve the multi-objective optimization function to obtain the Pareto optimal solution set, and the optimal combination of process parameters is determined according to production requirements. S6. When the state of the real-time grinding process changes, the multi-objective optimization function is dynamically updated to obtain the optimal combination of process parameters again.

2. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 1, characterized in that, Step S1, the preprocessing of the multi-source mixed signal, includes outlier detection and optimization procedures: First, the original signal is divided into multiple signal segments; Then, the signal segment is further divided into several signal parts. If the root mean square value of a certain signal part exceeds the set threshold, the signal part is regarded as an outlier. The threshold is derived according to the 3σ principle. Finally, outliers are removed from the original signal and the data is supplemented by linear interpolation.

3. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 1, characterized in that, The signal processing method based on WPD-EWT-FastICA in step S1 includes: The acoustic emission signal was processed using the WPD-FastICA method with a sampling frequency of 1MHz. Wavelet packet decomposition was used for three-level decomposition, and the db10 wavelet was selected as the wavelet basis function. The frequency band node signal with the highest energy proportion was selected from the decomposed node signal to construct the input matrix, and the independent components were separated by the FastICA algorithm. The vibration signal was processed using the EWT-FastICA method with a sampling frequency of 6400Hz. Empirical wavelet transform was used for decomposition, and the Fourier spectrum of the signal was adaptively segmented. The frequency band boundary was determined based on the spectral maxima. An empirical wavelet filter bank was constructed, and the decomposed components were used to construct an input matrix. The independent components were then separated using the FastICA algorithm. Independent components with peak frequencies around 300kHz are used as grinding burn source signal components, independent components with frequencies around 870Hz are used as grinding chatter source signal components, and independent components with peak frequencies around 450kHz are used as grinding wheel wear source signal components.

4. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 1, characterized in that, In step S3: Grinding chatter status labels include stable grinding status, slight chatter status, and severe chatter status; grinding wheel wear status includes initial wear status, normal wear status, and severe wear status; grinding burn status includes no burn status, slight burn status, and severe burn status.

5. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 4, characterized in that, In step S3: The method for classifying grinding chatter status labels is as follows: when the vibration signal spectrum energy distribution is uniform and there are no chatter marks on the workpiece surface, it is a stable grinding state; when the vibration signal spectrum energy begins to concentrate towards 870Hz and irregular grinding marks appear on the workpiece surface, it is a slight chatter state; when the vibration signal spectrum energy is concentrated at 870Hz and chatter marks form on the workpiece surface, it is a severe chatter state. The method for classifying the wear condition label of a grinding wheel is to detect the ratio of abrasive wear area R. area When 0 < R area When R is less than 16%, it is considered the initial wear state; when R is less than or equal to 16%, it is considered the initial wear state. area When R < 29%, it is considered normal wear; area When the wear rate is ≥ 29%, it is considered a severely worn condition; The method for classifying grinding burn status labels is as follows: test the surface Vickers hardness and the degree of oxidation discoloration. When HV ≥ 550 and the surface is bright yellow or has no discoloration, it is a no-burn status; when 470 < HV < 550 and the surface is yellowish-black or has slight discoloration, it is a slight burn status; when HV ≤ 470 and the surface is bluish-black and accompanied by cracks, it is a severe burn status.

6. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 5, characterized in that, The dynamic response mechanism in step S3 is as follows: If the output real-time grinding process status is slightly burned, severely burned, severely worn, or severely chattering, then the processing will be stopped. If the output real-time grinding process status is normal wear or slight chatter, the multi-objective optimization function is dynamically updated based on the current status and historical experimental and simulation data. The NSGA-III algorithm is used to solve the updated multi-objective optimization function and output the optimal combination of process parameters that fits the current status. When normal wear and slight chatter exist simultaneously, the dynamic update mechanism for slight chatter is given priority.

7. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 1, characterized in that, The recognition model based on CNN-LSTM-Attention consists of a convolutional neural network module, an attention mechanism module, and a long short-term memory network module. The convolutional neural network module is used to receive the input sensor time-series signal, which consists of data points at consecutive time steps. The convolutional kernel is slid along the time dimension through a one-dimensional convolution operation. The convolutional kernel covers multiple time steps to form local time windows, extracts the spatial features within each local time window, and outputs a feature sequence composed of time step features. The attention mechanism module is used to process the feature sequence, calculate the attention weight at each time step, and multiply the attention weight with the original feature at the corresponding time step element by element to enhance the feature representation of key time steps and generate a weighted feature sequence. The Long Short-Term Memory (LSTM) network module is used to receive weighted feature sequences, capture long-term dependencies between time steps, and output the final feature vector. The final feature vector is input into the Softmax classifier to obtain the state of the grinding process.

8. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 1, characterized in that, The multi-objective optimization function in step S5 includes: ， ， ， In the above three formulas, R a Roughness; σ is residual stress; Z w v represents the material removal rate per abrasive grain, used to calculate the total material removal rate (MRR). w v is the linear velocity of the workpiece. s a is the linear velocity of the grinding wheel; p This refers to the grinding depth. K x, y, z, a, b, c, d, e, f are obtained by fitting experimental data; τ is the ratio of the actual contact arc length to the moving contact arc length; k is the ratio of the actual grinding depth to the theoretical grinding depth; θ is the abrasive grain tip angle; d s d is the radius of the grinding wheel; w Let (v) be the radius of the workpiece; for (v) in the formula s ±v w When grinding against the grain, select +; when grinding with the grain, select -.

9. The method for dynamic optimization of surface quality in high-speed grinding of non-circular profiles of camshafts according to claim 1, characterized in that, The constraints in step S5 are: 。

Images