Method for predicting tool residual life based on data-physical fusion

Abstract

The application provides a kind of method for predicting tool residual life based on data-physical fusion, it is related to tool monitoring technical field, the application is by collecting milling cutter force signal, vibration signal and wear amount, pretreatment and feature extraction are carried out to obtain multidimensional feature matrix in time domain, frequency domain and time-frequency domain, and based on Mamba state space data-driven model is constructed, and the output tool wear prediction value is further established Tool wear physical model, analyze wear trend and determine wear critical point, statistical test of model parameters is carried out through historical data, form prior information, combine data-driven model with physical model, construct tool life data-physical collaborative fusion prediction model, particle filtering algorithm is applied to model optimization, finally predict the residual service life of tool.The application improves the accuracy of tool residual service life prediction by the fusion of data-driven and physical model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tool monitoring, in particular to a method for predicting tool residual life based on data-physical fusion. BACKGROUND

[0002] Tool residual service life prediction is an important link in improving production efficiency and reducing cost in modern manufacturing industry. In the field of mechanical processing, the state of the tool directly affects the processing quality, efficiency and cost. With the continuous development of manufacturing industry, the requirements for processing precision and production efficiency are getting higher and higher. It is crucial to timely and accurately grasp the wear state of the tool. The traditional tool life prediction method mainly depends on empirical formula and historical data. Although it can provide reference to a certain extent, it often cannot accurately reflect the complex environment and dynamic load influence on the tool in actual work, or it is monitored through space, but the space type is often single, and the actual state of the tool cannot be comprehensively and accurately reflected.

[0003] In the prior art, the tool life prediction in the traditional technology depends on a large amount of historical data and real-time monitoring data such as cutting force, cutting temperature, vibration signal and acoustic emission signal. Then, machine learning algorithm is used to analyze and process these data, and the mapping relationship between tool wear state and monitoring data is established. However, once the data is missing or inconsistent, the accuracy of the space will be affected, or the training data is insufficient or has noise, which will also affect the accuracy of the space. At the same time, the interpretability of the data-driven space is relatively poor, and it is difficult to intuitively understand how the space makes decisions.

[0004] In addition, the traditional physical space monitoring depends on the relationship between tool wear and physical parameters to realize the monitoring of tool wear state. The physical space can usually only consider a limited number of physical parameters. For complex processing environment and variable tool wear condition, the tool wear condition cannot be reflected in real time.

[0005] Therefore, it is necessary to provide a method for predicting tool residual life based on data-physical fusion to solve the above problems.

[0006] The above information disclosed in the background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0007] The present application aims to provide a method for predicting tool residual life based on data-physical fusion to solve the problems raised in the background.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0009] A method for predicting tool residual life based on data-physical fusion, the specific steps comprising:

[0010] Step 1: Collect the milling force signal, vibration signal and tool wear during the use of the tool, preprocess the collected signals, extract features from the preprocessed signals, obtain the features of the signals in time domain, frequency domain and time-frequency domain, and fuse the extracted signal features to obtain a multi-dimensional feature matrix of the tool;

[0011] Step 2: Taking the multi-dimensional feature matrix of the tool as input and the wear of the tool as output, establishing a tool data-driven model based on Mamba state space, and using a discrete parameter method to optimize and train the parameters of the tool data-driven model to output a tool wear prediction value;

[0012] Step 3: Establish a tool wear physical model, analyze the trend of tool wear, determine the wear critical point, and determine the model parameters according to the historical data, perform statistical test on the parameter distribution to determine the prior information of the parameters in the tool wear physical model;

[0013] Step 4: Combine the tool data-driven model and the tool wear physical model to establish a tool life data-physical collaborative fusion prediction model, optimize the tool life data-physical collaborative fusion prediction model using a particle filter algorithm, and predict the residual service life of the tool based on the optimized model.

[0014] Further, the collected milling force signal and vibration signal are preprocessed and feature extraction is performed to obtain a tool feature vector, and the method used is:

[0015] The milling force signal and vibration signal of the tool in X, Y and Z directions are obtained through a three-way force meter and an acceleration vibration sensor, the wear of the tool is calibrated through a Supereyes microscope, and the average wear of the four cutting edge relief of each milling cutter is taken as the wear of the tool;

[0016] The preprocessing operation of the collected sensor signal first removes the invalid signal collected during the feed and retreat of the original signal, removes the noise in the signal with a low-pass filter, inverts the negative half cycle of the signal for signal rectification, performs smoothing processing on the signal based on convolution operation, and finally normalizes the signal to scale the data to the interval [0, 1];

[0017] Feature extraction is performed from the preprocessed milling force signal and vibration signal, the time domain features include mean, variance and peak value, the frequency domain features include dominant frequency, spectral peak value and frequency band energy, and the time-frequency domain features include short-time Fourier transform energy, instantaneous frequency and wavelet energy;

[0018] The extracted time-domain, frequency-domain, and time-frequency-domain features are fused using the feature concatenation method to obtain time-domain feature vectors, frequency-domain feature vectors, and time-frequency-domain feature vectors, respectively. The three feature vectors are combined into a multidimensional feature matrix, and dimensionality reduction is performed using principal component analysis. Each column of the multidimensional feature matrix represents the usage parameters of a tool.

[0019] Furthermore, a tool data-driven model is established to output predicted tool wear values. The method used is as follows:

[0020] The tool data model based on the Mamba state space consists of state equations and observation equations, the expressions of which are as follows:

[0021] ;

[0022] ;

[0023] in, Indicates in The derivative of the hidden state at time t describes the rate of change of the hidden state over time. Indicates in A vector that hides the state at any given time, which typically includes tool state information. Let be the state transition matrix, matrix The elements are related to the physical properties and behavior of the cutting tool. The input matrix describes the effect of the input feature vectors on the changes in the hidden state. For tool data-driven models in The multidimensional feature matrix input at each time step. For tool data-driven models in The tool wear prediction value output at any time. Let be the observation matrix, representing the degree of contribution of different state components to the output. This is a direct influence matrix used to capture the linear effect of input on output;

[0024] The tool data-driven model is discretized, and the resulting equations are rewritten as follows:

[0025] ;

[0026] ;

[0027] in, For discrete time steps, represents the index in the discretized time series. Indicates at discrete time step The hidden state vector, Indicates at discrete time step The hidden state vector, Indicates at discrete time step The output, , Here, represents the discretized state transition matrix. It is the sampling time interval. , The input matrix is ​​discretized. It is the identity matrix. It is compensation for the impact of state transition;

[0028] A selection mechanism is introduced into the Mamba state space by setting the sampling time interval. , matrix , Represented as input The function is based on the following formula:

[0029] ;

[0030] in, , , All are selective space matrices. , , It is about choosing weights;

[0031] After discretization, a tool data-driven model is constructed using a selective state space, based on the following formula:

[0032] ;

[0033] in, Represents the selective state transition matrix. Represents the selective input matrix;

[0034] By employing selective state transitions and input matrices, the tool data-driven model can dynamically adjust the influence of data input on the hidden state. This is achieved by calculating the observation equations and obtaining the results at discrete time steps. The output indicates the time step. Predicted value of tool wear within the tool.

[0035] Furthermore, a physical model of tool wear is established, and the prior information of the parameters in the physical model of tool wear is determined. The method used is as follows:

[0036] Using the cutting time of the tool as input and the tool wear amount as output, a physical model of tool wear is established. This physical model consists of two stages: a convex stage and a concave stage. The convex stage represents the period from the start of tool wear to a certain point in time. , the wear of the tool increases with time in a logarithmic function form; the concave phase represents the process from the time point to the end of wear, in which the wear increases with time in a second-order polynomial form, and the description formulas of tool wear in the two phases are respectively set as:

[0037]

[0038]

[0039] wherein, represents the wear of the tool, , , , , are model parameters determined by regression analysis of historical data, represents the time of tool use;

[0040] In the above two description formulas of tool wear, there is a mixed area in the time period to , and this interval is applicable to both description formulas. By analyzing the trend of tool wear, the change in slope is reflected by using a first-order equation, and the first derivative is calculated to determine the wear critical point , thereby dividing the concave and convex two phases of tool wear, and the formula used is:

[0041]

[0042] wherein, represents the wear of the tool in the time period to , , are model parameters also determined by regression analysis of historical data, is the slope of the curve in the time period to ;

[0043] The wear data of the tool at different time points are collected, the values of and the corresponding wear of the tool are recorded, the historical data are fitted by using linear regression method to determine the specific values of , , , , , , , and is substituted into the first-order equation to obtain the wear critical point , and the tool wear is determined according to the following formula:​​​ To The secondary division of the mixing area of the time period is carried out, namely when , When , ;

[0044] Based on the constructed tool wear physical model, normal distribution, uniform distribution and exponential distribution are used to test whether the regression parameters conform to a specific distribution, and the P value of each parameter is recorded to determine its statistical significance. Based on the statistical test results of the parameters, the prior information suitable for the tool wear physical model is determined.

[0045] Further, a tool life data-physical collaborative fusion prediction model is established, and the method is as follows:

[0046] If the parameters in the tool wear physical model do not conform to the normal distribution, uniform distribution and exponential distribution, the tool wear prediction value output by the tool data-driven model is input into the tool wear physical model to obtain the model parameters, which are used as the prior information of the tool life data-physical collaborative fusion prediction model;

[0047] Integrate the tool wear prediction value output by the tool data-driven model and the prior information provided by the tool wear physical model, combine the physical model and the data-driven model, form a hierarchical fusion model structure, and the input of the fusion model is the use parameter of the tool, and the output is the residual service life prediction value of the tool. The tool life data-physical collaborative fusion prediction model first processes the input use parameter of the tool, predicts the wear state of the tool, and outputs the wear prediction value of the tool. According to the wear prediction value of the tool and the physical wear law of the tool, the residual life of the tool is predicted.

[0048] Further, the particle filtering algorithm is used to optimize the tool life data-physical collaborative fusion prediction model, and the optimized model is used to predict the residual service life of the tool, and the method is as follows:

[0049] Use the tool data-driven model and the tool wear physical model to provide observation values and prior information, and update the parameters in the tool life data-physical collaborative fusion prediction model in real time under the guidance of particle filtering. Based on the fact that tool wear is a continuous change process, a first-order autoregressive model is used to describe the tool wear state transition, and the formula is as follows:

[0050] ;

[0051] Where, represents the wear amount of the tool at time , is the state transition function at time , is process noise at time

[0052] The observation model for the tool wear state transition is:

[0053] ;

[0054] where, represents the tool wear observation value at time , is the observation function, is the observation noise;

[0055] At the initial time, randomly generate particles, each of which represents a tool remaining life. The parameter values obtained based on historical data are taken as the initial state of the particles, and each particle is assigned the same initial weight For each particle, its state at the current time is predicted according to the state transition, and the formula is:

[0056] ;

[0057] where, represents the tool wear of the th particle at time , represents the state transition function of the th particle at the current time , calculated from the state of the th particle at the previous time , is the process noise of the th particle at time

[0058] ;

[0059] ;

[0060] where, represents the weight of the th particle at time , represents the probability density function of the tool wear observation value under the condition that the particle state is , is the observation function, which is used to convert the particle state into the predicted tool wear observation value, is the variance of the observation noise;

[0061] In order to improve the diversity of particles, and further ensure the stability and convergence of the algorithm, the particles are resampled according to the weights of the particles to obtain a new set of particles, and the cumulative weight of each particle is calculated first, and the formula is:

[0062] ;

[0063] Wherein, represents the cumulative weight of the first particle, which is usually the starting weight in the particle filter, and the initial value is the weight of the first particle , is the cumulative weight of the particle , and a cumulative weight of each particle in the entire particle set is calculated, which represents the sum of the weights from the first particle to the particle , is the cumulative weight of the particle , is the weight of the particle , which is used to measure the matching degree of the particle and the observation data;

[0064] A random number is generated , a particle satisfying is found, which is copied to the new particle set, and finally the system resampling method is used until new particles are generated;

[0065] Based on the resampled particles, the estimated value of the tool wear is calculated, and the formula is:

[0066] ;

[0067] Wherein, represents the estimated value of the tool wear;

[0068] A maximum time step is set , and at each time step , the above steps are executed, including state prediction, weight update and system resampling, updating the particle and the system state, and after each time step ends, if , the iteration is terminated; otherwise, the next time step is executed;

[0069] The estimated value of the tool wear after the final iteration is completed is the predicted value of the remaining service life of the tool.

[0070] Compared with the prior art, the beneficial effects of the present application are:

[0071] The application carries out in-depth analysis on time series data of tool wear through Mamba state space, filters irrelevant information and extracts sensitive information, thereby establishing a data-driven model, providing high-quality observation data for a data-physical collaborative fusion prediction model, and in addition, according to the convex-concave characteristics of the tool wear curve, the tool wear curve is divided into two parts, and the corresponding physical model is established, the switching judgment of the formula is realized by analyzing the number of slope straight line data points, and finally, the data-driven model and the physical model are combined to provide observation data and prior information for the particle filtering algorithm, realize real-time dynamic prediction of the remaining service life of the tool, and optimize the tool management and maintenance strategy. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 It is the whole method flow diagram of the application.

[0073] Figure 2 It is the tool flank wear curve analysis diagram of the application. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical scheme and advantages of the application more clear and obvious, the application will be further described in detail below in combination with specific embodiments.

[0075] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the application should be understood as the usual meaning understood by those skilled in the art to which the application belongs. The "first", "second" and similar words used in the application do not represent any order, quantity or importance, but are only used to distinguish different components. "Include" or "contain" and similar words mean that the elements or objects before the word cover the elements or objects listed after the word and their equivalents, and do not exclude other elements or objects. "Connected" or "connected" and similar words are not limited to physical or mechanical connection, but can include electrical connection, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to represent the relative positional relationship, when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0076] EMBODIMENT:

[0077] Please refer to Figure 1 A method for predicting the remaining life of a tool based on data-physical fusion, the specific steps comprising:

[0078] Step 1: Collect the milling tool force signal, vibration signal and tool wear during the use of the tool, pretreat the collected signal, extract the features of the pretreated signal to obtain the features of the signal in time domain, frequency domain and time-frequency domain, and fuse the extracted signal features to obtain a multi-dimensional feature matrix of the tool.

[0079] In the specific embodiments of the present application, we carried out milling wear experiments, the experiments used VDL-1000E milling machine, the spindle speed was 2000r / min, the cutting Y depth was 0.4mm, the feed speed was 360mm / min, the Z axial cutting depth was 5mm, the workpiece material used in the milling process was 100mm*100mm*50mm titanium alloy TC4, and the cutting tool used was a 10mm flat head carbide end mill;

[0080] The three-direction force meter and the acceleration vibration sensor were used to obtain X, Y and Z three-direction signals, the signal sampling frequency was 10KHz, the collected sensor signals were preprocessed to remove the invalid signals collected during the feeding and retracting of the tool, and the collected force and vibration signals were truncated.

[0081] Further, the collected milling tool force signals and vibration signals were preprocessed and feature extraction was performed to obtain the feature vectors of the tool, and the method used was as follows:

[0082] The three-direction force meter and the acceleration vibration sensor were used to obtain the milling tool force signals and vibration signals of the tool in X, Y and Z three directions, the Supereyes microscope was used to calibrate the tool wear, and the average wear of the four cutting edges of each milling tool was taken as the tool wear;

[0083] The preprocessing operation of the collected sensor signals first removed the invalid signals collected during the feeding and retracting of the tool, used a low-pass filter to remove the noise in the signal, inverted the negative half cycle of the signal for signal rectification, performed smoothing processing on the signal based on convolution operation, and finally performed normalization processing on the signal to scale the data to the interval [0, 1];

[0084] Features were extracted from the preprocessed milling tool force signals and vibration signals, the time domain features included mean, variance and peak value, the frequency domain features included main frequency, spectral peak value and frequency band energy, and the time-frequency domain features included short-time Fourier transform energy, instantaneous frequency and wavelet energy;

[0085] The extracted time domain, frequency domain and time-frequency domain features were fused by feature splicing method to obtain time domain feature vectors, frequency domain feature vectors and time-frequency domain feature vectors, the three feature vectors were combined into a multi-dimensional feature matrix, and principal component analysis method was used for dimension reduction processing, and each list of the multi-dimensional feature matrix represented a tool use parameter.

[0086] Step 2: Using the multidimensional feature matrix of the tool as input and the tool wear amount as output, establish a tool data-driven model based on the Mamba state space, and use the discrete parameter method to optimize and train the parameters of the tool data-driven model to output the predicted value of tool wear amount.

[0087] In a specific embodiment of this invention, we use the Mamba state space to perform labeling and analysis on time series data, filtering irrelevant information and deeply mining sensitive information to establish a data-driven model. This provides observational information for the establishment of data-physical fusion. Firstly, continuous parameters are converted into discrete parameters through discretization to meet the requirements of machine learning. In the classic state-space model, the state equation and observation equation are the core of the state-space model.

[0088] We build our tool data-driven model based on the Mamba state space because it effectively models the multidimensional characteristics of tool wear, fully utilizing the dynamic nature of time series data for labeling analysis and sensitive information mining. This approach enables accurate descriptions of complex system states and improves model predictive performance by optimizing the training of discrete parameters. Furthermore, the Mamba state space supports the effective fusion of observational data with physical models, providing a solid foundation for real-time prediction of tool wear, ultimately contributing to more precise tool management and maintenance decisions.

[0089] Furthermore, a tool data-driven model is established to output predicted tool wear values. The method used is as follows:

[0090] The tool data model based on the Mamba state space consists of state equations and observation equations, the expressions of which are as follows:

[0091] ;

[0092] ;

[0093] in, Indicates in The derivative of the hidden state at time t describes the rate of change of the hidden state over time. Indicates in A vector that hides the state at any given time, which typically includes tool state information. Let be the state transition matrix, matrix The elements are related to the physical properties and behavior of the cutting tool. The input matrix describes the effect of the input feature vectors on the changes in the hidden state. For tool data-driven models in The multidimensional feature matrix input at each time step. For tool data-driven models in The tool wear prediction value output at any time. Let be the observation matrix, representing the degree of contribution of different state components to the output. This is a direct influence matrix used to capture the linear effect of input on output;

[0094] The tool data-driven model is discretized, and the resulting equations are rewritten as follows:

[0095] ;

[0096] ;

[0097] in, For discrete time steps, represents the index in the discretized time series. Indicates at discrete time step The hidden state vector, Indicates at discrete time step The hidden state vector, Indicates at discrete time step The output, , Here, represents the discretized state transition matrix. It is the sampling time interval. , The input matrix is ​​discretized. It is the identity matrix. It is compensation for the impact of state transition;

[0098] A selection mechanism is introduced into the Mamba state space by setting the sampling time interval. , matrix , Represented as input The function is based on the following formula:

[0099] ;

[0100] in, , , All are selective space matrices. , , It is about choosing weights;

[0101] After discretization, a tool data-driven model is constructed using a selective state space, based on the following formula:

[0102] ;

[0103] wherein, denotes a selective state transition matrix, denotes a selective input matrix;

[0104] By the selective state transition and input matrices, the tool data-driven model is able to dynamically adjust the influence of data input on hidden states, and by calculating the observation equation, the output at the discrete time step is obtained, which represents the tool wear prediction value within the time step .

[0105] Step 3: Establish a tool wear physical model, analyze the trend of tool wear, determine the wear critical point, and determine the model parameters according to historical data, and perform statistical test on the parameter distribution to determine the prior information of the parameters in the tool wear physical model.

[0106] In the specific embodiment of the present application, according to the trend change characteristics of the tool wear curve, it is divided into two parts of convex and concave, and the evolution process of the two parts is described by two formulas, and it is found that the formulas of the two parts of the wear curve have a mixed area in the time period to . This interval is suitable for both formulas describing the convex and concave. How to reasonably apply the formula model in this interval to accurately describe the wear process is a key point of the research content of this paper, and it can be seen from the straight line representing the slope that before the time period Figure 2 to , the slope of the tool wear curve is gradually getting smaller; after the time period to , the slope of the tool wear curve is gradually getting larger, so it can be inferred that in the mixed area of the time period to , there must be a turning point, which is the critical point of the slope change of the wear curve in the mixed area, so according to the analysis result of the above step, the present application takes the critical point as the judgment point of the wear process described by the two formulas, and uses the logarithmic function to describe the evolution process of the tool wear before it, and uses the second-order polynomial to describe the wear process after it. Then how to determine the specific position of the point becomes a new key point, and the change range of the angle of the straight line can directly show the size of the slope change. Therefore, this paper uses a linear equation to reflect the change of the slope, and obtains the first derivative to obtain the specific value of the slope.

[0107] ​The parameter distribution of two formulas in the physical model is analyzed, K2 test is used to test whether the parameters conform to normal distribution, and Kolmogorov-Smirnov test is used to test whether the parameters conform to uniform distribution and exponential distribution, through analysis, the P values of two parameters in the logarithmic function in the physical model are less than 0.05, which indicates that the parameters in the logarithmic function do not conform to normal distribution, uniform distribution and exponential distribution under the environmental conditions of the application, and the same test is conducted on the parameters of the second-order polynomial, and it is concluded that the three parameters in the second-order polynomial also do not conform to normal distribution, uniform distribution and exponential distribution;

[0108] According to the parameter distribution in the physical model, the prior condition can be set, if it conforms to which distribution, it can be directly set according to the distribution; in the present application, through analysis, the parameters in the physical model do not belong to any common distribution; considering that the particle filtering algorithm is used to simulate a large number of particles to approximate the posterior probability distribution, in this process, although accurate prior information can improve the performance and efficiency of the algorithm, even if the prior information is not accurate, the particle filtering algorithm can gradually correct and optimize the estimation of the posterior distribution through continuous iteration and updating, so the algorithm has relatively relaxed requirements for prior information; therefore, in the case that the current data does not meet the common distribution characteristics, the present application uses the tool wear prediction value output by the tool data driven model as input, so as to obtain the parameters in the physical model as prior information, it should be noted here that the tool wear used here is the flank wear of the milling cutter, since the flank wear directly affects the cutting performance and machining quality of the tool, the flank wear will cause the sharpness of the cutting edge to decrease, thereby affecting the cutting force, surface finish and machining accuracy, and the flank wear usually presents a relatively obvious wear pattern, such as wear band and wear point, which makes it easier to evaluate the wear degree of the tool through the measurement of the flank wear.

[0109] Further, a tool wear physical model is established, and the prior information of the parameters in the tool wear physical model is determined, and the method is as follows:

[0110] A tool wear physical model is established by taking the cutting time of the tool as input and taking the wear of the tool as output, the tool wear physical model has two stages, including a convex stage and a concave stage, the convex stage represents the process from the beginning of tool wear to a certain time point , in this stage, the wear of the tool increases in the form of a logarithmic function with the passage of time, the concave stage represents the process from the time point to the end of wear, in this stage, the wear increases in the form of a second-order polynomial with the passage of time, and the description formulas of tool wear in the two stages are respectively set as:

[0111] ;

[0112] ;

[0113] in, Indicates the amount of tool wear. , , , , These are all model parameters, determined through regression analysis of historical data. Indicates the time the tool was used;

[0114] In the two formulas describing tool wear above, it is in to There is a mixed region within the time period, and this interval is applicable to both descriptive formulas. By analyzing the trend of tool wear, a linear equation is used to reflect the slope change, and the first derivative is obtained to determine the wear critical point. This allows for the division of tool wear into two stages: unevenness and concavity. The formula used is as follows:

[0115] ;

[0116] in, express to The amount of tool wear over a period of time. , The model parameters were also determined through regression analysis of historical data. yes to The slope of the curve over a given time period;

[0117] Collect and record the wear data of the cutting tool at different time points. and the corresponding tool wear The value was determined by fitting historical data using linear regression. , , , , , , The specific value will Substitute into the linear equation to obtain the wear critical point. ,right to The time period is further divided into mixed regions, that is, when hour, ,when hour, ;

[0118] Based on the constructed tool wear physical model, normal distribution, uniform distribution and exponential distribution are used to test whether the regression parameters conform to a specific distribution, and the P value of each parameter is recorded to determine its statistical significance. Based on the statistical test results of the parameters, the prior information suitable for the tool wear physical model is determined.

[0119] Step 4: Combine the tool data-driven model with the tool wear physical model to establish a tool life data-physical collaborative fusion prediction model. Utilize the particle filtering algorithm to optimize the tool life data-physical collaborative fusion prediction model, and based on the optimized model, predict the remaining service life of the tool.

[0120] In the specific embodiments of the present application, we use data-driven models and physical models to provide observation values and prior information for the particle filtering algorithm, respectively. Under the guidance of particle filtering, the parameters in the physical model are updated in real time, thereby achieving accurate monitoring of tool wear. The reason for combining data-driven models with physical models is that data-driven models may be affected by noise and outliers, while the introduction of physical models can provide a more stable framework for data, enhancing the model's ability to adapt to uncertainty and changes. Subsequently, we use the particle filtering algorithm, and the model can be updated in real time when receiving new observation data, thereby reflecting changes in the tool wear state and ensuring the timeliness and accuracy of the prediction results.

[0121] Further, the tool life data-physical collaborative fusion prediction model is established by the following method:

[0122] If the parameters in the tool wear physical model do not conform to normal distribution, uniform distribution and exponential distribution, the tool wear prediction value output by the tool data-driven model is input into the tool wear physical model to obtain the model parameters, which are used as the prior information of the tool life data-physical collaborative fusion prediction model.

[0123] Integrate the tool wear prediction value output by the tool data-driven model with the prior information provided by the tool wear physical model. Combine the physical model and the data-driven model to form a hierarchical fusion model structure. The input of the fusion model is the tool usage parameters, and the output is the tool remaining service life prediction value. The tool life data-physical collaborative fusion prediction model first processes the input tool usage parameters, predicts the tool wear state, and outputs the tool wear prediction value. Then, based on the tool wear prediction value and the physical wear law of the tool, the remaining life of the tool is predicted.

[0124] It should be noted that the tool wear is a dynamic process, so we use particle filtering algorithm to update the model parameters in real time, adapt to the change of tool wear state, this adaptability ensures that the model always reflects the current wear condition, thereby improving the real-time and accuracy of the prediction, and through the first-order autoregressive model to describe the tool wear state transition and observation model combination, can better simulate the continuity and randomness of the tool wear process.

[0125] It should be noted that the randomly generated particles represent different possible tool remaining life, and the same initial weight is given to each particle to ensure fairness of the model at the beginning, and the weight of each particle is calculated based on the observation value, which can evaluate the matching degree with the actual wear amount, thereby improving the estimation accuracy of the wear state, in order to improve the diversity of particles and the stability of the algorithm, the resampling mechanism is introduced to effectively avoid the problem of particle degradation, by selecting particles that match the observation data better, the effectiveness of the particle set is maintained, thereby enhancing the reliability of the prediction results.

[0126] Therefore, it is necessary to optimize the tool life data-physical collaborative fusion prediction model by using particle filtering algorithm, and to predict the remaining useful life of the tool based on the optimized model, the method is as follows:

[0127] The tool data-driven model and the tool wear physical model are used to provide observation values and prior information, and the parameters in the tool life data-physical collaborative fusion prediction model are updated in real time under the guidance of particle filtering, based on the fact that tool wear is a continuous process, a first-order autoregressive model is used to describe the tool wear state transition, and the formula is as follows:

[0128] ;

[0129] Among them, represents the wear amount of the tool at time t, is the state transition function at time t, is the process noise at time t; The observation model for tool wear state transition is as follows:

[0130]

[0131] ; Among them,

[0132] represents the wear amount of the tool at time t, is the observation function, is the observation noise;

[0133] At the initial time, a plurality of particles are randomly generated​​​​ particles, each of which represents a tool residual life, the parameter values obtained based on historical data are taken as initial states of the particles, and each particle is given the same initial weight For each particle, its state at the current time is predicted according to state transition, and the formula is as follows:

[0134] ;

[0135] wherein, represents the wear amount of the tool at the time t of the i th particle, represents the state of the i th particle at the previous time t-1, the state transition function of the particle at the current time t is calculated, is the process noise of the i th particle at the time t; Under the observation model, the weight of each particle is calculated, and the formula is as follows:

[0136] ;

[0137] ;

[0138] wherein, represents the weight of the i th particle at the time t, represents the probability density function of the wear amount observation value of the tool under the condition that the state of the particle is known, is an observation function for converting the state of the particle into the predicted wear amount observation value of the tool, is the variance of the observation noise;

[0139] In order to improve the diversity of the particles and ensure the stability and convergence of the algorithm, the particles are resampled according to their weights to obtain a new set of particles. The cumulative weight of each particle is calculated first, and the formula is as follows:

[0140] ;

[0141] wherein, represents the cumulative weight of the first particle, which is usually the initial weight in particle filtering, and the initial value is the weight of the first particle , ​​​​​​​​​​​​The cumulative weight of each particle is calculated for every particle in the entire particle set, representing the weight from the first particle to the sixth particle. The sum of the weights of each particle. For the first The cumulative weight of each particle, For the first The weight of each particle is a value used to measure how well the particle matches the observed data;

[0142] Generate a random number Find a satisfactory The particles are copied to a new particle set, and finally, a system resampling method is used until the generated particles are obtained. A new particle;

[0143] The estimated value of tool wear is calculated based on the resampled particles, using the following formula:

[0144] ;

[0145] in, This represents an estimated value of tool wear.

[0146] Set a maximum time step. At each time step The above steps include state prediction, weight update, and system resampling, updating the particle and system states. At the end of each time step, if... If the iteration fails, the iteration terminates; otherwise, the iteration continues to the next time step.

[0147] Estimated tool wear after final iteration This is the predicted value of the remaining service life of the tool.

[0148] Please see Figure 2 In the schematic diagram of the tool flank wear curve analysis, the tool wear went through three stages: initial wear, normal wear, and severe wear. Figure 2 The cyan, orange, and red straight lines with dots at the beginning and end represent the slope of the relevant wear amount at that moment. The blue three-dimensional dots represent the critical points of the wear stage. Figure 2 As can be seen from the data, the entire wear curve can be divided into two parts. The first part is from the start of tool wear to the end of time. The shape of this wear curve can be considered as convex, and the process can be described by the following logarithmic function: The second part is from the perspective of time. The wear curve at the end of its wear cycle can be considered concave, and its wear process can be described by a second-order polynomial function: The two have a mixed region. To From Figure 2 It can be seen from To Before To After To In the mixed region There is a turning point in the mixed region, which is the critical point of the slope change of the wear curve, and the wear curve before the point uses a logarithmic function to represent the tool wear process, and the wear curve after the point

[0149] The above formulas are all dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain the most real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0150] The above embodiments can be realized wholly or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized in the form of a computer program product wholly or partially. Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized by hardware or software methods depends on the specific application and design constraints of the technical solutions.

[0151] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, which can be located in one place or distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0152] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for predicting the remaining tool life based on data-physical fusion, characterized in that, The specific steps include: Step 1: Collect the milling cutter force signal, vibration signal and tool wear during tool use, preprocess the collected signals, extract features from the preprocessed signals to obtain the features of the signals in the time domain, frequency domain and time-frequency domain, and fuse the extracted signal features to obtain the multidimensional feature matrix of the tool. Step 2: Using the multidimensional feature matrix of the tool as input and the tool wear amount as output, establish a tool data-driven model based on the Mamba state space, and use the discrete parameter method to optimize and train the parameters of the tool data-driven model to output the predicted value of tool wear amount. Step 3: Establish a physical model of tool wear, analyze the trend of tool wear, determine the wear critical point, and determine the model parameters based on historical data. Perform statistical tests on the parameter distribution to determine the prior information of the parameters in the physical model of tool wear. Step 4: Combine the tool data-driven model with the tool wear physical model to establish a tool life data-physical co-fusion prediction model. Use the particle filter algorithm to optimize the tool life data-physical co-fusion prediction model, and predict the remaining tool life based on the optimized model. The method used to establish the tool life data-physical co-fusion prediction model is as follows: Based on normal, uniform, and exponential distributions, prior information suitable for the physical model of tool wear is determined. The predicted tool wear value output by the tool data-driven model is integrated with the prior information provided by the physical model of tool wear. The physical model and the data-driven model are combined to form a hierarchical fusion model structure. The input of the fusion model is the tool's usage parameters, and the output is the predicted value of the tool's remaining service life. The tool life data-physical co-fusion prediction model first processes the input tool usage parameters, predicts the tool's wear state, and outputs the predicted tool wear value. Then, based on the predicted tool wear value and the physical wear law of the tool, the remaining tool life is predicted.

2. The method for predicting remaining tool life based on data-physical fusion according to claim 1, characterized in that, The acquired milling cutter force and vibration signals are preprocessed and feature extracted to obtain the tool's feature vector. The method used is as follows: The milling cutter force and vibration signals in the X, Y, and Z directions are obtained by using a three-dimensional force gauge and an acceleration vibration sensor. The wear of the cutter is calibrated by using a Supereyes microscope. The average wear of the back face of the four cutting edges of each milling cutter is taken as the wear of the cutter. The preprocessing operation of the acquired sensor signal first removes the invalid signals acquired during tool entry and exit from the original signal, removes noise from the signal using a low-pass filter, reverses the negative half-cycle of the signal for signal rectification, smooths the signal based on convolution operation, and finally normalizes the signal to scale the data to the interval [0,1]. Feature extraction is performed from the preprocessed milling cutter force signal and vibration signal. The time domain features include mean, variance and peak value; the frequency domain features include main frequency, spectral peak value and frequency band energy; and the time-frequency domain features include short-time Fourier transform energy, instantaneous frequency and wavelet energy. The extracted time-domain, frequency-domain, and time-frequency-domain features are fused using the feature concatenation method to obtain time-domain feature vectors, frequency-domain feature vectors, and time-frequency-domain feature vectors, respectively. The three feature vectors are combined into a multidimensional feature matrix, and dimensionality reduction is performed using principal component analysis. Each column of the multidimensional feature matrix represents the usage parameters of a tool.

3. The method for predicting remaining tool life based on data-physical fusion according to claim 2, characterized in that, A tool data-driven model is established to output predicted tool wear values. The method used is as follows: The tool data model based on the Mamba state space consists of state equations and observation equations, the expressions of which are as follows: ； ； in, Indicates in The derivative of the hidden state at time t describes the rate of change of the hidden state over time. Indicates in The vector of hidden states at any given time includes the tool's state information. Let be the state transition matrix, matrix The elements are related to the physical properties and behavior of the cutting tool. The input matrix describes the effect of the input feature vectors on the changes in the hidden state. For tool data-driven models in The multidimensional feature matrix input at each time step. For tool data-driven models in The tool wear prediction value output at any time. Let be the observation matrix, representing the degree of contribution of different state components to the output. This is a direct influence matrix used to capture the linear effect of input on output; The tool data-driven model is discretized, and the resulting equations are rewritten as follows: ； ； in, For discrete time steps, represents the index in the discretized time series. Indicates at discrete time step The hidden state vector, Indicates at discrete time step The hidden state vector, Indicates at discrete time step The output, , Here, represents the discretized state transition matrix. It is the sampling time interval. , The input matrix is ​​discretized. It is the identity matrix. It is compensation for the impact of state transition; A selection mechanism is introduced into the Mamba state space by setting the sampling time interval. , matrix , Represented as input The function is based on the following formula: ； in, , , All are selective space matrices. , , It is about choosing weights; After discretization, a tool data-driven model is constructed using a selective state space, based on the following formula: ； in, Represents the selective state transition matrix. Represents the selective input matrix; By employing selective state transitions and input matrices, the tool data-driven model can dynamically adjust the influence of data input on the hidden state. This is achieved by calculating the observation equations and obtaining the results at discrete time steps. The output indicates the time step. Predicted value of tool wear within the tool.

4. The method for predicting remaining tool life based on data-physical fusion according to claim 3, characterized in that, A physical model of tool wear was established, and the prior information of the parameters in the physical model of tool wear was determined. The method used was as follows: Using the cutting time of the tool as input and the tool wear amount as output, a physical model of tool wear is established. This physical model consists of two stages: a convex stage and a concave stage. The convex stage represents the period from the start of tool wear to a certain point in time. During this stage, the wear of the tool increases logarithmically over time; The concave stage represents the time point During the process leading to the end of wear, the wear amount increases over time in the form of a second-order polynomial. The descriptive formulas for tool wear in the two stages are set as follows: ； ； in, Indicates the amount of tool wear. , , , , These are all model parameters, determined through regression analysis of historical data. Indicates the time the tool was used; In the above two stages of tool wear description formulas, it is in to There is a mixed region within the time period, and this interval is applicable to both descriptive formulas. By analyzing the trend of tool wear, a linear equation is used to reflect the slope change, and the first derivative is obtained to determine the wear critical point. This allows for the division of tool wear into two stages: unevenness and concavity. The formula used is as follows: ； in, express to The amount of tool wear over a period of time. , The model parameters were also determined through regression analysis of historical data. yes to The slope of the curve over a given time period; Collect and record the wear data of the cutting tool at different time points. and the corresponding tool wear The value was determined by fitting historical data using linear regression. , , , , , , The specific value will Substitute into the linear equation to obtain the wear critical point. ,right to The time period is further divided into mixed regions, that is, when hour, ,when hour, .

5. The method for predicting remaining tool life based on data-physical fusion according to claim 4, characterized in that, The method used to determine prior information is as follows: Based on the constructed physical model of tool wear, normal, uniform and exponential distributions are used to test whether the regression parameters follow a specific distribution. The p-value of each parameter is recorded to determine its statistical significance. Based on the statistical test results of the parameters, the prior information applicable to the physical model of tool wear is determined. If the parameters in the tool wear physical model do not follow a normal, uniform, or exponential distribution, then the predicted tool wear value output by the tool data-driven model is input into the tool wear physical model to obtain model parameters, which serve as prior information for the tool life data-physical co-fusion prediction model.

6. The method for predicting remaining tool life based on data-physical fusion according to claim 5, characterized in that, The particle filtering algorithm is used to optimize the tool life data-physical co-fusion prediction model. Based on the optimized model, the remaining tool life is predicted. The method is as follows: This paper utilizes tool data-driven models and tool wear physical models to provide observations and prior information. Guided by particle filtering, the parameters in the tool life data-physical co-fusion prediction model are updated in real time. Based on the premise that tool wear is a continuously changing process, a first-order autoregressive model is used to describe the transition of tool wear states. The underlying formula is: ； in, Indicates in The amount of tool wear at any given time. Is The state transition function at time t, yes Process noise at time step; The observation model on which the wear state transition of the cutting tool is based is: ； in, Indicates in The observed value of tool wear at time [time]. It is the observation function. It is observation noise; Randomly generated at the initial time. Each particle represents a type of tool's remaining life. Parameter values ​​obtained from historical data are used as the initial state of the particles, and each particle is assigned the same initial weight. For each particle, the formula used to predict its state at the current moment based on the state transition is: ； in, Indicates the first Individual particles The amount of tool wear at any given time. Indicates from the first The particles in the previous moment status Calculate the particle's position at the current moment. State transition function, It is the first Individual particles Process noise at time step; Under the observation model, the weight of each particle is calculated using the following formula: ； in, Indicates the first Individual particles The weight of the particle at time t. Indicates that the known particle state is Under these circumstances, the observed value of tool wear The probability density function, It is an observation function used to determine the state of a particle. Converted into predicted tool wear observations. The variance of the observed noise; To improve particle diversity and thus ensure the stability and convergence of the algorithm, the particles are resampled based on their weights to obtain a new set of particles. The cumulative weight of each particle is first calculated using the following formula: ； in, This represents the cumulative weight of the first particle, which is the initial weight in particle filtering. The initial value is the weight of the first particle. , For the first The cumulative weight of each particle is calculated for every particle in the entire particle set, representing the weight from the first particle to the sixth particle. The sum of the weights of each particle. For the first The cumulative weight of each particle, For the first The weight of each particle is a value used to measure how well the particle matches the observed data; Generate a random number Find a satisfactory The particles are copied to a new particle set, and finally, a system resampling method is used until the generated particles are obtained. A new particle; The estimated value of tool wear is calculated based on the resampled particles, using the following formula: ； in, This represents an estimated value of tool wear. Set a maximum time step. At each time step The above steps include state prediction, weight update, and system resampling, updating the particle and system states. At the end of each time step, if... If the iteration fails, the iteration terminates; otherwise, the iteration continues to the next time step. Estimated tool wear after final iteration This is the predicted value of the remaining service life of the tool.

Images