CN115186571A - Method for predicting residual service life of cutter under different working conditions based on Bi-GRU network - Google Patents
Method for predicting residual service life of cutter under different working conditions based on Bi-GRU network Download PDFInfo
- Publication number
- CN115186571A CN115186571A CN202210525232.3A CN202210525232A CN115186571A CN 115186571 A CN115186571 A CN 115186571A CN 202210525232 A CN202210525232 A CN 202210525232A CN 115186571 A CN115186571 A CN 115186571A
- Authority
- CN
- China
- Prior art keywords
- service life
- cutter
- residual service
- working condition
- under different
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
- B23Q—DETAILS, COMPONENTS, OR ACCESSORIES FOR MACHINE TOOLS, e.g. ARRANGEMENTS FOR COPYING OR CONTROLLING; MACHINE TOOLS IN GENERAL CHARACTERISED BY THE CONSTRUCTION OF PARTICULAR DETAILS OR COMPONENTS; COMBINATIONS OR ASSOCIATIONS OF METAL-WORKING MACHINES, NOT DIRECTED TO A PARTICULAR RESULT
- B23Q17/00—Arrangements for observing, indicating or measuring on machine tools
- B23Q17/09—Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool
- B23Q17/0995—Tool life management
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention provides a method for predicting the residual service life of a cutter under different working conditions based on a Bi-GRU network. According to the method, the increment thought of a regularization joint training mode is adopted, force signals in the milling process are collected through the field processing process, time domain and frequency domain analysis is carried out on the collected signals, relevant features are extracted, feature screening is carried out by adopting a Maximum Information Coefficient (MIC) method, and screened features are fused by combining a kernel function principal component analysis (KPCA) method, so that the purpose of reducing the dimension is achieved, and the condition of the residual service life of a cutter in the milling process is researched.
Description
Technical Field
The invention relates to a method for predicting the residual service life of a cutter under different working conditions based on a Bi-GRU network.
Background
In numerical control machining, the quality of machining quality directly influences the performance of parts, and in the real machining process, a used cutter inevitably directly influences the machining quality of products along with performance degradation. At present, in an actual production workshop, whether a tool is changed or not is usually determined by adopting a mode of regular tool changing or artificial subjective judgment, and the problems that the residual service life of a numerical control tool cannot be identified, is difficult to predict and the like exist. The utilization rate of the cutter is reduced by changing the cutter too early, the manufacturing cost is improved, and serious consequences such as product surface quality out-of-tolerance, cutter damage, even casualties and the like can be caused by changing the cutter too late. In order to ensure the processing quality of products and improve the utilization rate of the numerical control cutter, the residual service life of the numerical control cutter needs to be accurately predicted so as to accurately grasp the health condition and the potential fault condition of the cutter, thereby ensuring safe and stable production and processing.
The accurate prediction of the residual service life of the cutter has important significance for ensuring the quality of the machined surface and reducing the loss caused by cutter degradation. In the production and manufacturing of the fields of aerospace ships and the like, various and small-batch processing modes become typical characteristics of the fields, the numerical control cutter can experience different processing working conditions in the actual use process, the residual service life of the numerical control cutter under different processing working conditions can be changed due to the adjustment of factors such as workpiece material properties, workpiece structural characteristics, actual cutting parameters, cutter geometric parameters, machine tool self characteristics and the like in the cutting process, the factors are mutually coupled and have a nonlinear incidence relation, so that once the use working conditions of the cutter are changed, if a large amount of historical data marking cutter degradation are obtained again, time and labor are consumed. The original prediction model trained under the original working condition is difficult to be applied to the data sample under the new working condition, so that the reuse effect of the prediction model is weakened, the prediction precision is reduced, and finally the performance of the prediction model is invalid. In order to learn new knowledge, the traditional machine learning algorithm can only abandon the existing model, re-analyze the problem and train the prediction model from the initial state, so that repeated learning of historical data is caused, a large amount of time, space, manpower and material resources are wasted, and the residual service life of the numerical control cutter under different working conditions is difficult to predict.
Disclosure of Invention
Technical problem to be solved
How to design a generalized prediction model to accurately judge the residual service life of the numerical control cutter under different working conditions has important practical significance. The following problems are faced to realize accurate prediction of the remaining service life of the numerical control cutter under different working conditions in the production process:
(1) Multiple working condition factors, dynamic change and strong coupling
In the actual production process, products in the military industry such as aerospace ships belong to a small-batch production mode, the residual service life of the numerical control cutter is coupled by various working condition factors, the action mechanism relationship among the factors is complex and fuzzy, the process scene and the processing object of the numerical control cutter are variable, the residual service life of the cutter is mutually influenced by the working condition factors such as part materials, processing technologies and workpiece structures and the processing process signal factors such as bending moment signals and torque signals, so that the problems have the characteristics of numerous types of influencing factors, dynamic changes and mutual coupling, and a clear and universal mathematical model is difficult to construct to accurately predict the residual service life of the numerical control cutter.
(2) The signal data of the processing process has large fluctuation, complex types and high redundancy
Numerical control machining is a complex system engineering in actual production, time-varying data such as bending moment signals, torque signals and the like in the machining process are large in fluctuation, complex in type and high in redundancy, and are simultaneously influenced by multiple factors such as artificial subjective factors and workshop machining environments, data noise conditions such as data loss, data repetition and data errors exist in a used high-precision machine tool and a used data acquisition system, so that the relation of monitoring signals in the machining process is difficult to mine and cannot be directly used for model input.
(3) The change of the residual service life of the numerical control cutter has time sequence relevance
During actual milling, the remaining service life of the numerical control cutter is related to the current machining working condition and the historical machining working condition, namely, the change condition of the previous moment can have different influence on the change condition of the later moment, so that the value change process of the remaining service life of the numerical control cutter can be followed by a certain rule, independent fracture research can not be carried out, and the time sequence relevance is realized. Therefore, the built numerical control tool residual service life prediction model needs to have the capability of processing the time sequence correlation.
(4) Generalization difference of residual service life prediction model of numerical control cutter
In the field of military industry such as aerospace ships and the like, the numerical control cutter used for production has wide actual application scenes, the residual service life of the numerical control cutter changes along with the change of factors such as processing parameters, machine tool precision, workpiece materials, process characteristics and the like, a common prediction model is limited to the life prediction under a specific working condition, and the numerical control cutter under a new working condition is likely to have the condition that the effect of the prediction model is weakened or even fails.
Technical scheme
Aiming at the technical problem, the invention provides a method for predicting the residual service life of a cutter under different working conditions based on a bidirectional gated circulation network (Bi-GRU) network. According to the method, the increment thought of a regularization joint training mode is adopted, force signals in the milling process are collected through the field processing process, time domain and frequency domain analysis is carried out on the collected signals, relevant features are extracted, feature screening is carried out by adopting a Maximum Information Coefficient (MIC) method, and screened features are fused by combining a kernel function principal component analysis (KPCA) method, so that the purpose of reducing the dimension is achieved, and the condition of the residual service life of a cutter in the milling process is researched.
The prediction models under different working conditions must have the capabilities of training, learning and processing the newly added samples, and meanwhile, the accurate prediction capability of the original samples is kept. Namely, the method has two meanings: one refers to the ability to learn new knowledge and the other refers to the inability to corrupt previously learned knowledge. If the parameters are too many and the model is too complex, overfitting is easily caused, namely the model performs well on training sample data, but performs poorly on an actual test sample and does not have good generalization capability. In order to avoid overfitting, the L2 regularization method is used, the square sum of the weight parameters is added on the basis of the original loss function, the limiting parameters are too much or too large, the model is prevented from being more complex, the prediction under different working conditions is realized, and the requirements of actual engineering are met.
The technical scheme of the invention is as follows:
the method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network comprises the following steps of:
step1: collecting monitoring signal data of a numerical control cutter in the real machining process, uniformly representing the factors of the current working condition, and recording a residual service life label life and a machining quality constraint condition corresponding to the working condition;
step2: preprocessing the monitoring signal data adopted in the step 1; performing time domain feature extraction and frequency domain feature extraction on the preprocessed monitoring signal data; then, carrying out characteristic correlation analysis on the extracted characteristics and the residual service life of the cutter to obtain strong correlation characteristics which meet requirements on the correlation with the residual service life of the cutter; then, carrying out feature fusion and dimension reduction on the obtained strong correlation features to obtain a feature set sensitive to the service life of the tool;
and step3: establishing a Bi-GRU network model, and aiming at the current working condition, taking the current working condition factors uniformly represented in the step1, the feature set sensitive to the service life of the cutter obtained in the step2, the current state of the cutter and the residual service life label corresponding to the working condition as model training samples, and establishing a regression learner to train the Bi-GRU network model to obtain a numerical control cutter residual service life prediction model f under the working condition;
and 4, step 4: when the prediction model f fails to work in the face of a new working condition task, the original prediction model f is subjected to fine adjustment by adopting an L2 regularized incremental learning joint training mode to obtain a super model f' which can predict both new task data and original task data, and the prediction of the numerical control cutter RUL under different working conditions is realized.
Further, in step1, the monitoring signal data of the numerical control tool in the real machining process are torque, bending moment in the X direction and bending moment in the Y direction.
Further, in step1, the working condition factors include a process parameter sub-working condition P, a workpiece information sub-working condition W, a machine tool information sub-working condition O, a cutting fluid property sub-working condition F, and a cutter information sub-working condition Cut.
Further, the process of uniformly characterizing the working condition factors is as follows:
a process parameter sub-condition P, expressed as P = [ a = [) p n f way]Wherein a is p The cutting depth is shown, n is the rotating speed of the main shaft, f is the feed amount, and way is the feed mode; the feed mode way is represented by one-hot coding;
the workpiece information sub-condition W is represented as W = [ K E mu s tau Rm zeta HRA Ak R ], wherein K represents the thermal conductivity coefficient of the material, E represents the Young modulus of the material, mu s represents the friction coefficient, tau represents the Poisson ratio, rm represents the tensile strength, zeta represents the shear strength, HRA represents the Rockwell hardness, ak represents the impact toughness, and R represents the melting point of the material;
machine tool information sub-condition O, expressed as O = [ T = [ ] in T out Sys Axle]Wherein T is in Indicating the internal ambient temperature, T, of the machine tool out The external environment temperature of the machine tool is represented, sys represents the machining precision of the machine tool, and Axle represents the number of main shafts of the machine tool;
the cutting fluid property sub-condition F is expressed as F = [ lambda PH T F v]Wherein λ represents the conductivity of the cutting fluid, PH represents the pH value of the cutting fluid, and T F Represents the temperature of the cutting fluid, v represents the ejection speed of the cutting fluid;
the tool information sub-condition Cut, expressed as Cut = [ z D γ L a ], where z represents the number of teeth of the tool, D represents the tool diameter, β represents the helix angle degree, L represents the tool overhang, and a represents the tool type.
Further, in step1, the remaining service life of the numerical control tool is represented by the machining stroke, and the machining quality constraint condition is represented by the surface roughness Ra of the machined workpiece.
Further, in step2, the process of preprocessing the monitoring signal data includes: intercepting an effective value, processing a missing value, standardizing data and processing Kalman filtering.
Further, in step2, the time domain features extracted from the preprocessed monitoring signal data include a mean, a standard deviation, a skewness, a kurtosis, a root mean square and a form factor; the extracted frequency domain features comprise a power spectrum mean value, a power spectrum root mean square, a power spectrum crest factor, a power spectrum stability ratio and a power spectrum improvement equivalent bandwidth.
Further, in step2, a maximum information coefficient method is adopted to analyze the correlation between the extracted features and the residual service life of the cutter, so as to obtain strong correlation features meeting the requirements on the correlation between the extracted features and the residual service life of the cutter.
Further, in step2, a kernel function principal component analysis method is adopted to perform feature fusion dimension reduction on the obtained strong correlation features to obtain a feature set sensitive to the service life of the tool.
Further, in step 4, the incremental learning joint training mode using L2 regularization refers to: and introducing a weight attenuation norm in the original loss function by adopting an L2 regularization method.
Advantageous effects
By analyzing the problem of the service life of the cutter in the numerical control machining process and aiming at the complex coupling relation between the working condition factors and the cutter performance degradation in the milling machining process, the invention provides the method for predicting the residual service life of the numerical control cutter under different working conditions, which is based on the unified characterization of the working condition factors, takes the extraction of the characteristics of the monitoring signals as the condition, takes the incremental learning as the idea and takes the neural network as the core.
Aiming at the problems of redundant repetition, strong noise loss and the like of data in numerical control processing, technologies such as effective value interception, missing value processing, standardization, kalman Filtering (KF) and the like are utilized for preprocessing, so that the quality of original data is improved to a great extent; in order to reduce the number of characteristics of monitoring signals, time domain analysis, frequency domain analysis and other characteristics of the processed data are extracted, and the characteristics are screened and fused by combining a Maximum Information Coefficient (MIC) method and a kernel function principal component analysis (KPCA) method, so that the problems of information redundancy or loss caused by excessive characteristics and random selection are solved, and the generalization of a numerical control tool residual service life prediction model is ensured.
Aiming at the problems that time accumulation effect exists between working condition factors and cutter performance degradation in the machining process and the residual service life of the cutter can be influenced to a certain extent by the future working conditions, the idea of incremental learning is adopted, the working condition factors related to the residual service life of the cutter are used as the input of a prediction model, an L2 regularized incremental learning combined training mode is adopted, a Bi-GRU network is used for learning the newly added working conditions, the new working conditions are continuously fused and learned to obtain a more accurate prediction model, and the method has accurate prediction effect on the residual service life of the cutter.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1: a flow chart of the prediction step;
FIG. 2: the overall structure of the tool remaining service life prediction method.
Detailed Description
In the production and manufacturing of the fields of aerospace ships and the like, various and small-batch machining modes become typical characteristics of the fields, the numerical control cutter can experience different machining working conditions in the actual use process, under different machining working conditions, the residual service life of the numerical control cutter can be changed due to the adjustment of workpiece material characteristics, workpiece structural characteristics, actual cutting parameters, cutter geometric parameters and machine tool characteristics, the factors are coupled with one another and have a nonlinear incidence relation, and therefore once the machining working conditions change, if a large amount of historical data for marking cutter degradation are obtained again, time and labor are consumed. The original prediction model under the original working condition is difficult to apply to the data sample under the new working condition, so that the reuse effect of the prediction model is weakened, the prediction precision is reduced, and finally the performance of the prediction model is invalid. In order to learn new knowledge, the traditional machine learning algorithm can only abandon the existing model, re-analyze the problem and train the prediction model from the initial state, so that repeated learning of historical data is caused, a large amount of time, space, manpower and material resources are wasted, and the residual service life of the numerical control cutter under different working conditions is difficult to predict.
The embodiment provides a method for predicting the residual service life of a cutter under different working conditions based on a Bi-GRU network, which specifically comprises the following steps:
step1: collecting monitoring signal data of a numerical control cutter in a real machining process, wherein the monitoring signal data comprises torque (Mon-torque), X-direction bending moment (Mon-bending (X)) and Y-direction bending moment (Mon-bending (Y))), uniformly representing current working condition factors, and recording a residual service life label life and a machining quality constraint condition corresponding to the working condition.
In the real machining process, factors influencing the residual service life of the numerical control cutter are various, and a generalizable working condition model is difficult to form, so that the effect of the prediction model under a new working condition is weakened, and even the prediction model fails. In contrast, the method and the device represent the working condition factors under different working conditions in the form of the working condition vector, and avoid weakening the generalization of the prediction model due to the diversification of the working condition factors. The working condition factors comprise a process parameter sub-working condition P, a workpiece information sub-working condition W, a machine tool information sub-working condition O, a cutting fluid property sub-working condition F and a cutter information sub-working condition Cut.
Process parameter sub-condition P:
during the machining process, different technological parameters can cause the change of cutting force, so that the real-time health state of the numerical control cutter is influenced, and the prediction of the residual service life is influenced. The invention mainly summarizes the process parameter sub-working conditions as cutting depth and main working conditionsShaft speed, feed amount and feed mode: expressed as P = [ a ] p n f way]Wherein a is p Representing the cutting depth, n representing the rotating speed of a main shaft, f representing the feed amount, and way representing the feed mode; the feed way is represented by one-hot coding.
Workpiece information sub-condition W:
the processed workpiece is used as an object which is in close contact with the numerical control cutter, and the information of the processed workpiece is a key factor influencing the stability of the processing process and an important factor influencing the state change of the numerical control cutter. The invention takes the main parameters of heat conductivity coefficient, young modulus, friction coefficient, poisson's ratio, tensile strength, shearing strength, rockwell hardness, impact toughness, melting point and the like as consideration factors to carry out formalized characterization: expressed as W = [ K E μ s τ Rm ζ HRA Ak R ], where K represents the thermal conductivity of the material, E represents the young's modulus of the material, μ s represents the coefficient of friction, τ represents the poisson's ratio, rm represents the tensile strength, ζ represents the shear strength, HRA represents the rockwell hardness, ak represents the impact toughness, and R represents the melting point of the material.
Machine tool information sub-condition O:
in general, the self state of machine tool equipment inevitably affects the machining process, and when predicting the remaining service life of the numerical control tool, the machine tool information is taken as an essential consideration, but the slight influence caused by the degradation and damage of the machine tool is not considered, specifically, O = [ T ]) in T out Sys Axle]Wherein T is in Indicating the internal ambient temperature, T, of the machine tool out The external environment temperature of the machine tool is shown, sys shows the machining accuracy of the machine tool, and Axle shows the number of spindles of the machine tool.
Cutting fluid property sub-condition F:
in the numerical control machining process, the cutting fluid quality, the cutting heat and the cutting force have a certain correlation relationship, and the self state of the cutting tool is further influenced, specifically expressed as F = [ lambda PH T ] F v]Wherein λ represents the conductivity of the cutting fluid, PH represents the pH value of the cutting fluid, and T F The temperature of the cutting fluid is shown, and v is the injection speed of the cutting fluid.
Tool information sub-condition Cut:
the invention discloses a numerical control cutter health state evaluation method, which is characterized in that the self attribute of a cutter is taken as a decisive factor of the residual service life of the numerical control cutter, and has important significance in cutter health state evaluation.
And 2, step: with the rapid development of sensor technology, data acquisition in the actual processing process is more efficient, and because the data acquired in the processing process has the characteristic of massive redundancy, the raw data needs to be preprocessed, extracted, screened and fused, so that the information loss caused by overhigh feature dimension and random selection of features input by a prediction model is avoided, the generalization of the prediction model is weakened, and the accurate prediction of the residual service life of the numerical control cutter under different working conditions is difficult to realize.
Therefore, the monitoring signal data adopted in the step1 is preprocessed; performing time domain feature extraction and frequency domain feature extraction on the preprocessed monitoring signal data; then, carrying out feature correlation analysis on the extracted features and the residual service life of the cutter to obtain strong correlation features meeting requirements on the correlation between the extracted features and the residual service life of the cutter; and then carrying out feature fusion dimension reduction on the obtained strong correlation features to obtain a feature set sensitive to the service life of the tool.
The monitoring signal data acquired in the real machining process of the numerical control cutter covers the processes of the cutter from the initial idle stroke stage, the workpiece contact stage, the complete cutting-in stage, the stable cutting stage, the cutter cutting-out stage, the cutter retracting idle stroke stage and the like in the machining process, the capacity of a data signal segment is huge, the acquired signal data has the characteristics of large data volume, low value density, different sources, high redundancy and the like due to the complex cutting process of the actual machining, the defects weaken the data quality to different degrees and reduce the reliability of the numerical control cutter, and the premise of accurate and reliable prediction of the RUL is the authenticity and integrity of the data, and the data with the unqualified quality can cause overlarge deviation of a prediction result after being input into a model, so that the prediction effect is seriously influenced. Therefore, preliminary preprocessing of the raw data is necessary at an early stage, including: intercepting an effective value, processing a missing value, standardizing data and processing Kalman filtering.
(1) Interception of effective value
The method comprises the steps of collecting complete machining data signals of the numerical control cutter from a new cutter to a scrapped full life cycle, wherein the complete machining data signals comprise the conditions of shutdown, idle cutter feeding, cutter changing, island avoidance and the like, wherein the condition that the cutter is not in contact with a workpiece exists inevitably, and the data obtained under the condition have no essential significance for predicting the residual service life of the numerical control cutter. Therefore, it is necessary to appropriately intercept the entire signal data, thereby improving the processing efficiency. The invention takes the boundary points of the cutting force data mutation as the upper and lower limits of the intercepted range, and takes the effective data of the stable cutting stage for analysis.
(2) Missing value handling
When data is acquired in a real processing scene, the acquired data is usually lost due to some uncontrollable factors (such as degradation of a machine tool, sudden change of environment, sensitivity of an acquisition system and the like), and according to a coding mechanism inside a computer, a lost value is generally coded into a space, naN or other placeholders, and the lost value causes the conditions of low model accuracy, error report of a prediction program and the like in a subsequent use process of the data. In this embodiment, a layered mean interpolation is used to complete missing values in the monitoring signal. The hierarchical mean interpolation accurately levels the variables according to the attribute characteristics of the variables before interpolation, ensures that the characteristics of data of each layer are similar, and takes the mean value of a complete unit in each layer of data as an interpolation value of the layer. The processing effect of the layered mean interpolation method is obvious, the data after complete supplementation is reasonable, and the distribution deviation of the data and the whole data is not overlarge, so that the subsequent calculation and analysis are seriously influenced.
(3) Data normalization
Data acquired in the machining process generally have different orders of magnitude and dimensions, when level differences under different working conditions are large, the fact that a certain factor occupies too large weight in a model due to the fact that the numerical order is too large can weaken the weight of key factors with great importance but small numerical orders in the prediction model, and the final effect of the prediction model is seriously affected, so that the original data need to be subjected to standardization processing, the influence of the factors with the dimensions and the orders of magnitude on subsequent calculation and analysis is eliminated to the maximum extent, and the convergence speed and the prediction accuracy of the model are further guaranteed. In the embodiment, a Z-score method is adopted for standardization, dimensions and magnitude levels are uniformly represented, the inherent property of original data is not lost, and valuable information is reserved.
(4) Kalman filtering process
The monitoring signal obtained by the intelligent tool handle contains more random noise due to unstable factors of the actual production environment, and the key for preprocessing the monitoring signal is to ensure that the key information of the data is mined, so that the Kalman filtering is adopted to update and process the data acquired on site in real time.
Aiming at the problems that the sampling points of the preprocessed signals are dense and too high in frequency, and the subsequent prediction model is difficult to directly input, the embodiment performs time domain feature extraction and frequency domain feature extraction on the preprocessed monitoring signal data:
the time domain feature extracted from the preprocessed monitoring signal data comprises:
x in the table i (i =1,2, \8230;, n) represents a sequence of sample points of the original signal.
The extracted frequency domain features include:
wherein f is i Signal sequence x representing cutting force i A corresponding frequency; p i Represents f i A corresponding power spectrum;
represents f i Is measured.
And (3) extracting 7 types of time domain features and 5 types of frequency domain features, and totaling 12 types of time domain and frequency domain features, wherein the composition of the cutting force signal comprises 3 dimensions of information including torque (Mon-torque), X-direction bending moment (Mon-bending (X)) and Y-direction bending moment (Mon-bending (Y)), so as to obtain a feature matrix of 12X 3 =36. For a subsequent numerical control tool residual service life prediction model, the prediction performance of the model is influenced by overhigh input characteristic dimension, so that further screening and optimization are needed.
And (3) performing characteristic correlation analysis on the extracted characteristics and the residual service life of the cutter by adopting a Maximum Information Coefficient (MIC) method to obtain strong correlation characteristics which meet requirements on the correlation with the residual service life of the cutter:
step1: monitoring signal data characteristic set { F extracted by time domain and frequency domain analysis i Total 36 features { F } 1 、F 2 、…、F 36 } and the remaining tool life { life } i Forming a new matrix D;
step2: distributing all data points in D on a two-dimensional plane by utilizing MIC analysis, performing grid division on the plane by using a multiplied by b straight lines, and calculating the probability distribution of each data point falling on a defined grid, wherein the probability distribution is called as 'mutual information' of F and life under the division scheme, and the formula is as follows:
step3: changing a grid division scheme to obtain different 'mutual Information', normalizing the 'mutual Information' to [0,1], comparing the sizes of the mutual Information obtained under grids of different dimensions, taking the maximum value MAX (Multual Information), and determining the maximum Information coefficient between the feature set and the residual service life of the cutter, wherein the maximum Information coefficient is shown as the following formula:
in the formula: b is the capacity function and B (n) = n α (ii) a n is the number of experimental samples; alpha is a factor coefficient influencing the universality of the MIC method, and alpha is 0.6.
Performing redundancy and correlation analysis on the extracted total 36 features, wherein the process of the change of the remaining service life of the cutter can be reflected by the features with higher correlation with the life features in the extracted features, and the obtained result is as follows:
for comparison, the pearson correlation coefficient method (PPMCC) is also used to perform characteristic correlation analysis on the extracted characteristics and the remaining service life of the tool, and the result is:
it can be seen that the correlation r = Corr (F, Y) absolute value of 9 features and the residual service life of the tool obtained by the PPMCC analysis method is greater than 0.6, that is, there are 9 strong correlation features; the MIC analysis method finds the correlation MIC [ F of 14 characteristics and the residual service life of the cutter; life is greater than 0.6 in absolute value, namely 14 strong correlation features exist, 9 features are consistent with those obtained by the PPMCC analysis method, and the other 5 features do not show linear correlation with the residual service life of the tool, but contain valuable non-linear correlation information. Therefore, if the PPMCC analysis method is used, the nonlinear related information is inevitably greatly lost, and the valuable and effective characteristic information can be comprehensively obtained by adopting the MIC analysis method.
The acquired monitoring information is an important basis for reflecting the health state of the numerical control cutter and is also a key analysis factor of the remaining service life of the cutter. Due to the gradual increase of the capacity of the collected information, the characteristics of the monitoring signals are increased, so that the characteristic information redundancy phenomenon is caused, and if partial characteristics are directly and randomly selected, the value of the monitoring information is lost. Therefore, the method adopts a kernel function principal component analysis method to perform feature fusion dimensionality reduction on the obtained strong correlation features to obtain a feature set sensitive to the service life of the tool, so that the value of original information is ensured to the maximum extent, the intrinsic information of data is more comprehensively represented, and the weakening of the generalization of a prediction model caused by excessive feature number or randomly selected features is avoided.
And step3: establishing a Bi-GRU network model, and aiming at the current working condition, taking the current working condition factors uniformly represented in the step1, the feature set sensitive to the service life of the cutter obtained in the step2, the current state of the cutter and the residual service life label corresponding to the working condition as model training samples, and establishing a regression learner to train the Bi-GRU network model to obtain a residual service life prediction model f of the numerical control cutter under the working condition;
and 4, step 4: when a new working condition task fails, the prediction model f adopts an L2 regularized increment learning joint training mode, introduces a weight attenuation norm into an original loss function, and finely adjusts the original prediction model f to obtain a 'super model f' which can predict both new task data and original task data, thereby realizing the prediction of the numerical control cutter RUL under different working conditions.
The invention uses the Bi-GRU network with increment learning ability for predicting the residual service life of the cutter under different working conditions, and gradually updates the network parameters in the learning process. The Bi-GRU network based on the L2 regularization joint training mode does not affect the original learned knowledge after learning a new sample in the new knowledge, and the prediction error is obviously reduced, namely the prediction accuracy of the network is improved to more than 94% after learning new data.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made in the above embodiments by those of ordinary skill in the art without departing from the principle and spirit of the present invention.
Claims (10)
1. A method for predicting the residual service life of a cutter under different working conditions based on a Bi-GRU network is characterized by comprising the following steps: the method comprises the following steps:
step1: collecting monitoring signal data of a numerical control cutter in a real machining process, uniformly representing current working condition factors, and recording a residual service life label life and a machining quality constraint condition corresponding to the working condition;
step2: preprocessing the monitoring signal data adopted in the step 1; performing time domain feature extraction and frequency domain feature extraction on the preprocessed monitoring signal data; then, carrying out characteristic correlation analysis on the extracted characteristics and the residual service life of the cutter to obtain strong correlation characteristics which meet requirements on the correlation with the residual service life of the cutter; then, performing feature fusion dimensionality reduction on the obtained strong correlation features to obtain a feature set sensitive to the service life of the cutter;
and 3, step3: establishing a Bi-GRU network model, and aiming at the current working condition, taking the current working condition factors uniformly represented in the step1, the feature set sensitive to the service life of the cutter obtained in the step2, the current state of the cutter and the residual service life label corresponding to the working condition as model training samples, and establishing a regression learner to train the Bi-GRU network model to obtain a numerical control cutter residual service life prediction model f under the working condition;
and 4, step 4: when the prediction model f fails to work in the face of a new working condition task, the original prediction model f is subjected to fine adjustment by adopting an L2 regularized incremental learning joint training mode to obtain a super model f' which can predict both new task data and original task data, and the prediction of the numerical control cutter RUL under different working conditions is realized.
2. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1, wherein the method comprises the following steps: in the step1, the monitoring signal data of the numerical control tool in the real machining process are torque, X-direction bending moment and Y-direction bending moment.
3. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1 or 2, wherein the method comprises the following steps: in the step1, the working condition factors comprise a process parameter sub-working condition P, a workpiece information sub-working condition W, a machine tool information sub-working condition O, a cutting fluid property sub-working condition F and a cutter information sub-working condition Cut.
4. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 3, wherein the method comprises the following steps: the process of uniformly characterizing the working condition factors comprises the following steps:
a process parameter sub-condition P, expressed as P = [ a = [) p n f way]Wherein a is p Representing the cutting depth, n representing the rotating speed of a main shaft, f representing the feed amount, and way representing the feed mode; the feed mode way is represented by one-hot coding;
the workpiece information sub-condition W is represented as W = [ K E mu s tau Rm zeta HRA Ak R ], wherein K represents the thermal conductivity of the material, E represents the Young modulus of the material, mu s represents the friction coefficient, tau represents Poisson ratio, rm represents tensile strength, zeta represents shear strength, HRA represents Rockwell hardness, ak represents impact toughness, and R represents the melting point of the material;
machine tool info sub-condition O, expressed as O = [ T = in T out Sys Axle]Wherein T is in Indicating the internal ambient temperature, T, of the machine tool out The external environment temperature of the machine tool is represented, sys represents the machining precision of the machine tool, and Axle represents the number of main shafts of the machine tool;
the cutting fluid property sub-condition F is expressed as F = [ lambda PH T F v]Wherein λ represents the conductivity of the cutting fluid, PH represents the pH value of the cutting fluid, and T F Represents the temperature of the cutting fluid, and v represents the injection speed of the cutting fluid;
the tool information sub-condition Cut, expressed as Cut = [ z D β L a ], where z represents the number of tool teeth, D represents the tool diameter, β represents the helix angle degree, L represents the tool overhang, and a represents the tool type.
5. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1, wherein the method comprises the following steps: in the step1, the remaining service life of the numerical control cutter is represented by the machining stroke, and the machining quality constraint condition is represented by the surface roughness Ra of the machined workpiece.
6. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1, wherein the method comprises the following steps: in step2, the process of preprocessing the monitoring signal data comprises the following steps: intercepting an effective value, processing a missing value, standardizing data and processing Kalman filtering.
7. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1 or 6, wherein the method comprises the following steps: in step2, extracting time domain characteristics of the preprocessed monitoring signal data, wherein the time domain characteristics comprise a mean value, a standard deviation, a skewness, a kurtosis, a root mean square and a form factor; the extracted frequency domain features comprise a power spectrum mean value, a power spectrum root mean square, a power spectrum crest factor, a power spectrum stability ratio and a power spectrum improvement equivalent bandwidth.
8. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1, wherein the method comprises the following steps: and 2, performing characteristic correlation analysis on the extracted characteristics and the residual service life of the cutter by adopting a maximum information coefficient method to obtain strong correlation characteristics which meet requirements on the correlation between the extracted characteristics and the residual service life of the cutter.
9. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1, wherein the method comprises the following steps: and 2, performing feature fusion dimensionality reduction on the obtained strong correlation features by adopting a kernel function principal component analysis method to obtain a feature set sensitive to the service life of the cutter.
10. The method for predicting the residual service life of the cutter under different working conditions based on the Bi-GRU network as claimed in claim 1, wherein the method comprises the following steps: in step 4, the incremental learning joint training mode using L2 regularization refers to: and introducing a weight attenuation norm in the original loss function by adopting an L2 regularization method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210525232.3A CN115186571A (en) | 2022-05-15 | 2022-05-15 | Method for predicting residual service life of cutter under different working conditions based on Bi-GRU network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210525232.3A CN115186571A (en) | 2022-05-15 | 2022-05-15 | Method for predicting residual service life of cutter under different working conditions based on Bi-GRU network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115186571A true CN115186571A (en) | 2022-10-14 |
Family
ID=83512676
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210525232.3A Pending CN115186571A (en) | 2022-05-15 | 2022-05-15 | Method for predicting residual service life of cutter under different working conditions based on Bi-GRU network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115186571A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115687984A (en) * | 2023-01-05 | 2023-02-03 | 绍兴市特种设备检测院 | Method for monitoring health state of stirring kettle |
| CN116673793A (en) * | 2023-08-03 | 2023-09-01 | 比亚迪股份有限公司 | Tool loss detection method, medium, electronic device and tool loss detection device |
| CN116738868A (en) * | 2023-08-16 | 2023-09-12 | 青岛中德智能技术研究院 | Rolling bearing residual life prediction method |
| CN117592976A (en) * | 2024-01-19 | 2024-02-23 | 山东豪泉软件技术有限公司 | Cutter residual life prediction method, device, equipment and medium |
-
2022
- 2022-05-15 CN CN202210525232.3A patent/CN115186571A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115687984A (en) * | 2023-01-05 | 2023-02-03 | 绍兴市特种设备检测院 | Method for monitoring health state of stirring kettle |
| CN116673793A (en) * | 2023-08-03 | 2023-09-01 | 比亚迪股份有限公司 | Tool loss detection method, medium, electronic device and tool loss detection device |
| CN116673793B (en) * | 2023-08-03 | 2023-11-14 | 比亚迪股份有限公司 | Tool loss detection method, medium, electronic device and tool loss detection device |
| CN116738868A (en) * | 2023-08-16 | 2023-09-12 | 青岛中德智能技术研究院 | Rolling bearing residual life prediction method |
| CN116738868B (en) * | 2023-08-16 | 2023-11-21 | 青岛中德智能技术研究院 | Rolling bearing residual life prediction method |
| CN117592976A (en) * | 2024-01-19 | 2024-02-23 | 山东豪泉软件技术有限公司 | Cutter residual life prediction method, device, equipment and medium |
| CN117592976B (en) * | 2024-01-19 | 2024-04-26 | 山东豪泉软件技术有限公司 | Cutter residual life prediction method, device, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115186571A (en) | Method for predicting residual service life of cutter under different working conditions based on Bi-GRU network | |
| Shi et al. | Tool wear prediction via multidimensional stacked sparse autoencoders with feature fusion | |
| US11761930B2 (en) | Prediction method of part surface roughness and tool wear based on multi-task learning | |
| Dou et al. | An unsupervised online monitoring method for tool wear using a sparse auto-encoder | |
| Abou-El-Hossein et al. | Prediction of cutting force in end-milling operation of modified AISI P20 tool steel | |
| CN103488135B (en) | A kind of statistical process control method for semiconductor production machining process monitoring | |
| CN113496312A (en) | Cutter residual life prediction method and equipment based on multi-dimensional feature extraction fusion and long-term and short-term memory network and storage medium | |
| CN107627152A (en) | A kind of digital control processing chip control method based on BP neural network | |
| CN113780153A (en) | Cutter wear monitoring and predicting method | |
| CN112101439B (en) | High-speed wire quality defect diagnosis and tracing method based on distributed Bayesian network | |
| Yao et al. | Modelling of multivariate time series for tool wear estimation in finish-turning | |
| Li et al. | Multiconditional machining process quality prediction using deep transfer learning network | |
| CN113627544A (en) | Machine tool milling cutter state identification method based on multi-source heterogeneous data fusion | |
| Zuperl et al. | Intelligent cutting tool condition monitoring in milling | |
| Zhang et al. | Optimization of machining parameters of 2.25 Cr1Mo0. 25V steel based on response surface method and genetic algorithm | |
| Varatharajulu et al. | Experimental investigation and multi objective optimisation of Duplex 2304 drilling operation using evolutionary algorithm | |
| Wu et al. | Tool wear assessment and life prediction model based on image processing and deep learning | |
| Jiang et al. | A dynamic failure mode and effect analysis (FMEA) method for CNC machine tool in service | |
| CN114357851A (en) | Numerical control milling surface roughness prediction method based on DAE-RNN | |
| Al Masud et al. | Modeling of Chip Tool Interface Temperature in Machining Steel-An Artificial Intelligence (AI) Approach | |
| Cui et al. | Modeling of dimensional errors in slender bar turning using artificial neural networks | |
| Zhang et al. | Prediction of Tool Remaining Useful Life of NC Machine Tool Based on DTW Algorithm and LSTM Neural Network | |
| Miao et al. | Research on Operation Status Prediction of Production Equipment Based on Digital Twins and Multidimensional Time Series | |
| Tao et al. | Health state evaluation for fuzzy multi-state production systems based on MPNM | |
| Liu et al. | Prediction of cutting force via machine learning: state of the art, challenges and potentials |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |