A Prediction Method for the Remaining Service Life of Tools under Different Working Conditions Based on Bi-GRU Network

Abstract

The present invention proposes a method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network. By borrowing the incremental idea of the regularization joint training method, the method collects the force signal during the milling process through on-site machining, analyzes the collected signal in the time domain and frequency domain, extracts relevant features, uses the maximum information coefficient method (MIC) for feature screening, and then combines the kernel principal component analysis (KPCA) to fuse the screened features to achieve the purpose of dimensionality reduction, so as to study the remaining service life of the tool during the milling process.

Description

Technical Field

[0001] The present invention relates to a method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network. Background Art

[0002] In numerical control machining, the quality of machining directly affects the performance of parts. In the actual machining process, the tool used will inevitably directly affect the machining quality as its performance deteriorates. Currently, in actual production workshops, the decision of whether to change the tool is usually made by regular tool changing or manual subjective judgment. There are often problems such as the inability to identify and difficult to predict the remaining service life of numerical control tools. Premature tool changing reduces the tool utilization rate and increases the manufacturing cost. Delayed tool changing will lead to serious consequences such as poor surface quality of the product, tool damage, and even personal injury. To ensure the machining quality of products and improve the utilization rate of numerical control tools, it is necessary to accurately predict the remaining service life of numerical control tools to accurately grasp the tool health status and potential failure conditions, thereby ensuring safe and stable production machining.

[0003] Accurately predicting the remaining service life of a tool is of great significance for ensuring the machining surface quality and reducing the losses caused by tool degradation. In the production and manufacturing of fields such as aerospace and shipbuilding, the processing mode of multiple varieties and small batches has become a typical feature of this field. Numerical control tools will experience different machining working conditions during actual use. During the cutting process, the remaining service life of numerical control tools under different machining working conditions will change due to the adjustment of factors such as workpiece material properties, workpiece structure characteristics, actual cutting parameters, tool geometric parameters, and the characteristics of the machine tool itself. These factors are coupled with each other and have a non-linear correlation relationship. Once the tool working conditions change, if a large amount of historical data marking tool degradation is obtained again, it will be very time-consuming and laborious. The original prediction model trained under the original working conditions is difficult to apply to the data samples under the new working conditions, resulting in a weakened reuse effect and reduced prediction accuracy of the prediction model, and ultimately leading to the performance failure of the prediction model. To learn new knowledge, traditional machine learning algorithms can only abandon the existing model, re-analyze the problem, and train the prediction model from the initial state, resulting in repeated learning of historical data, wasting a large amount of time, space, manpower, and material resources. This makes it difficult to predict the remaining service life of numerical control tools under different working conditions. Summary of the Invention

[0004] Technical Problems to be Solved

[0005] How to design a more generalized prediction model to accurately judge the remaining service life of numerical control tools under different working conditions has important practical significance. To accurately predict the remaining service life of numerical control tools under different working conditions during the production process, the following difficulties are faced:

[0006] (1) There are many types of operating factors, dynamic changes, and strong coupling

[0007] In the actual production process, products in the military industry such as aerospace and shipbuilding belong to the small-batch production mode. The remaining service life of CNC tools is affected by the coupling of multiple working conditions, and the mechanism relationship between each factor is complex and vague. The process scenarios and processing objects of CNC tools change unpredictably. The remaining service life of the tool affects working conditions such as part materials, processing technology, workpiece structure, and processing process signal factors such as bending moment signals and torque signals. As a result, this type of problem has the characteristics of many types of influencing factors, dynamic changes, and mutual coupling. It is difficult to construct a clear and universal mathematical model to accurately predict the remaining service life of CNC tools.

[0008] (2) The signal data of the processing process is highly volatile, complex in type and highly redundant

[0009] CNC machining itself is a complex system engineering in actual production. The time-varying data such as bending moment signals and torque signals in the machining process are highly volatile, complex in type, and highly redundant. They are also affected by multiple factors such as human subjective factors and the machining environment of the workshop. In addition, the high-precision machine tools and data acquisition systems used will have data noise such as missing data, duplications, and errors. As a result, the relationship between the monitoring signals in the machining process is difficult to explore and cannot be directly used for model input.

[0010] (3) The change of the remaining service life of CNC tools has a time series correlation

[0011] In actual milling processing, the remaining service life of CNC tools is related to both the current processing conditions and the historical processing conditions, that is, the changes in the previous moment will have different effects on the changes in the next moment, which makes the change process of the remaining service life of CNC tools follow certain rules and cannot be studied separately, and has time series correlation. Therefore, the prediction model for the remaining service life of CNC tools established needs to have the ability to handle time series correlation.

[0012] (4) The generalization of the remaining service life prediction model of CNC tools is poor

[0013] In the military industry such as aerospace and shipbuilding, the actual application scenarios of CNC tools used in production are extensive, and their remaining service life varies with factors such as processing parameters, machine tool accuracy, workpiece materials and process characteristics. Common prediction models are only limited to life prediction under specific working conditions. When facing new working conditions, the prediction model effect of CNC tools may be weakened or even fail.

[0014] Technical Solution

[0015] To solve the above technical problems, the present invention proposes a method for predicting the remaining useful life of a tool under different working conditions based on a bidirectional gated recurrent unit (Bi-GRU) network. By borrowing the incremental idea of the regularization joint training method, the method collects the force signals during the milling process through on-site machining, analyzes the collected signals in the time domain and frequency domain, extracts relevant features, uses the maximum information coefficient method (MIC) for feature screening, and then combines the kernel principal component analysis (KPCA) to fuse the screened features to achieve the purpose of dimensionality reduction, so as to study the remaining useful life of the tool during the milling process.

[0016] The prediction models under different working conditions must have the ability to train, learn, and process new samples, and at the same time maintain the accurate prediction ability for the original samples. That is, it has two meanings: one refers to the ability to learn new knowledge, and the other refers to not destroying the knowledge learned before. If there are too many parameters and the model is too complex, it is easy to cause overfitting, that is, the model performs well on the training sample data but poorly on the actual test samples, and does not have good generalization ability. To avoid overfitting, the present invention uses the L2 regularization method, adding the sum of the squares of the weight parameters to the original loss function, restricting the parameters from being too many or too large, avoiding the model from being more complex, realizing the prediction under different working conditions, and meeting the requirements of actual engineering.

[0017] The technical solution of the present invention is as follows:

[0018] The method for predicting the remaining useful life of a tool under different working conditions based on a Bi-GRU network includes the following steps:

[0019] Step 1: Collect the monitoring signal data during the actual machining process of the numerical control tool, uniformly characterize the current working condition factors, and record the remaining useful life label life and the machining quality constraint conditions corresponding to this working condition;

[0020] Step 2: Preprocess the monitoring signal data used in Step 1; extract the time-domain features and frequency-domain features of the preprocessed monitoring signal data; then perform feature correlation analysis on the extracted features and the remaining useful life of the tool to obtain strongly correlated features whose correlation with the remaining useful life of the tool meets the requirements; then fuse and reduce the dimensionality of the obtained strongly correlated features to obtain a feature set sensitive to the tool life;

[0021] Step 3: Establish a Bi-GRU network model. For the current working condition, use the current working condition factors uniformly characterized in Step 1, the feature set sensitive to the tool life obtained in Step 2, the current state of the tool, and the remaining useful life label corresponding to this working condition as the model training samples, and construct a regression learner to train the Bi-GRU network model to obtain the prediction model f of the remaining useful life of the numerical control tool under this working condition;

[0022] Step 4: When the prediction model f fails to handle a new working condition task, an incremental learning joint training method with L2 regularization is adopted to fine-tune the original prediction model f, obtaining a "super model f'" that can predict both new task data and original task data, thus realizing the prediction of the remaining useful life (RUL) of the numerical control tool under different working conditions.

[0023] Furthermore, in Step 1, the monitoring signal data during the actual machining process of the numerical control tool are torque, bending moment in the X direction, and bending moment in the Y direction.

[0024] Furthermore, in Step 1, the working condition factors include the process parameter sub-condition P, workpiece information sub-condition W, machine tool information sub-condition O, cutting fluid property sub-condition F, and tool information sub-condition Cut.

[0025] Furthermore, the process of uniformly characterizing the working condition factors is as follows:

[0026] The process parameter sub-condition P is expressed as P = [a p n f way], where a p represents the cutting depth, n represents the spindle speed, f is the feed rate, and way represents the feed mode; the feed mode way is represented by one-hot encoding;

[0027] The workpiece information sub-condition W is expressed as W = [K E μs τ Rm ζ HRA Ak R], where K represents the thermal conductivity of the material, E represents the Young's modulus of the material, μs represents the friction coefficient, 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;

[0028] The machine tool information sub-condition O is expressed as O = [T in T out Sys Axle], where T in represents the internal environment temperature of the machine tool, T out represents the external environment temperature of the machine tool, Sys represents the machining accuracy of the machine tool, and Axle represents the number of spindles of the machine tool;

[0029] The cutting fluid property sub-condition F is expressed as F = [λ PH T F v], where λ represents the conductivity of the cutting fluid, PH represents the pH value of the cutting fluid, T F represents the cutting fluid temperature, and v represents the injection speed of the cutting fluid;

[0030] The cutting tool information sub - working condition Cut is expressed as Cut = [z D γ L A], where z represents the number of cutting tool teeth, D represents the cutting tool diameter, β represents the spiral angle degree, L represents the cutting tool overhang, and A represents the cutting tool type.

[0031] Furthermore, in step 1, the remaining service life of the CNC cutting tool is characterized by the machinable stroke, and the machining quality constraint condition is characterized by the surface roughness Ra of the workpiece to be machined.

[0032] Furthermore, in step 2, the process of pre - processing the monitoring signal data includes: effective value interception, missing value processing, data standardization, and Kalman filtering processing.

[0033] Furthermore, in step 2, the time - domain features extracted from the pre - processed monitoring signal data include mean, standard deviation, skewness, kurtosis, peak factor, root mean square, and waveform factor; the frequency - domain features extracted include power spectrum mean, power spectrum root mean square, power spectrum peak factor, power spectrum stability ratio, and power spectrum improved equivalent bandwidth.

[0034] Furthermore, in step 2, the maximum information coefficient method is used to perform feature correlation analysis between the extracted features and the remaining service life of the cutting tool, and strong - correlation features whose correlation with the remaining service life of the cutting tool meets the requirements are obtained.

[0035] Furthermore, in step 2, the kernel principal component analysis method is used to perform feature fusion and dimensionality reduction on the obtained strong - correlation features to obtain a feature set sensitive to the cutting tool life.

[0036] Furthermore, in step 4, the incremental learning joint training method using L2 regularization means: using the L2 regularization method to introduce a weight decay norm into the original loss function.

[0037] Beneficial effects

[0038] By analyzing the cutting tool life problem in the CNC machining process, aiming at the complex coupling relationship between the working condition factors and the degradation of cutting tool performance in milling machining, the present invention proposes a prediction method for the remaining service life of CNC cutting tools under different working conditions, which is based on the unified characterization of working condition factors, conditional on the extraction of monitoring signal features, with the idea of incremental learning, and with the neural network as the core.

[0039] Aiming at problems such as redundant repetition and strong noise and missing values in data presentation during numerical control machining, preprocessing is carried out using techniques such as valid value interception, missing value processing, standardization, and Kalman filtering (KF), which greatly improves the quality of the original data. To reduce the number of features of the monitoring signal, time-domain analysis, frequency-domain analysis, and other feature extractions are performed on the processed data, and the screening and fusion of features are realized by combining the maximum information coefficient method (MIC) and kernel principal component analysis (KPCA), avoiding the problems of information redundancy or loss caused by excessive features and random selection, and ensuring the generalization of the prediction model for the remaining service life of the numerical control tool.

[0040] Aiming at the problem that there is a time accumulation effect between the working condition factors and the tool performance degradation during the machining process, and the future working conditions will also have a certain impact on the remaining service life of the tool, the idea of incremental learning is adopted. The working condition factors related to the remaining service life of the tool are used as the input of the prediction model. The incremental learning joint training method with L2 regularization is adopted, and the Bi-GRU network is used to learn the new working conditions. The new working conditions are continuously fused and learned to obtain a more accurate prediction model, which has an accurate prediction effect on the remaining service life of the tool. Compared with the existing methods under specific single working conditions, this method has good generalization performance and can accurately predict the remaining service life of the tool under different working conditions, which is of great significance for realizing the application of the prediction of the remaining service life of the tool in actual production.

[0041] The additional aspects and advantages of the present invention will be partly given in the following description, partly will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above and / or additional aspects and advantages of the present invention will become obvious and easy to understand from the description of the embodiments in conjunction with the following drawings, wherein:

[0043] Figure 1 : Flow chart of the prediction steps;

[0044] Figure 2 : Overall architecture of the method for predicting the remaining service life of the tool. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] In the production and manufacturing of fields such as aerospace and ships, the processing mode of multi-variety and small-batch has become a typical feature of this field. During the actual use of CNC tools, they will go through different processing conditions. Under different processing conditions, the remaining service life of CNC tools will change due to the characteristics of workpiece materials, the structural features of workpieces, actual cutting parameters, tool geometric parameters, and the adjustment of machine tool characteristics. These factors are coupled with each other and have a non-linear correlation. Therefore, once the processing conditions change, it will be very time-consuming and laborious to obtain a large amount of historical data marking tool degradation again. The original prediction model under the original working conditions is difficult to apply to the data samples under the new working conditions, resulting in a weakened reuse effect of the prediction model and a reduced prediction accuracy, and ultimately leading to the performance failure of the prediction model. In order to learn new knowledge, traditional machine learning algorithms can only abandon the existing model, re-analyze the problem, and train the prediction model from the initial state, resulting in repeated learning of historical data, wasting a large amount of time, space, manpower and material resources, which makes it difficult to predict the remaining service life of CNC tools under different working conditions.

[0046] In this embodiment, a method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network is proposed, which specifically includes the following steps:

[0047] Step 1: Collect the monitoring signal data during the actual machining process of the CNC tool, including torque (Mon-torque), bending moment in the X direction (Mon-bending(x)), and bending moment in the Y direction (Mon-bending(y))), and uniformly characterize the current working condition factors, and record the remaining service life label life and the machining quality constraint condition corresponding to this working condition. In this embodiment, the remaining service life of the CNC tool is characterized by the machinable travel, and the machining quality constraint condition is characterized by the surface roughness Ra of the workpiece to be machined.

[0048] During the actual machining process, there are many factors affecting the remaining service life of the CNC tool, and it is difficult to form a generalizable working condition model, resulting in a weakened effect of the prediction model under the new working condition or even failure. In response to this, the present invention represents the working condition factors under different working conditions in the form of a working condition vector to avoid the weakening of the generalization of the prediction model caused by the diversification of working condition factors. The working condition factors include the process parameter sub-working condition P, the workpiece information sub-working condition W, the machine tool information sub-working condition O, the cutting fluid attribute sub-working condition F, and the tool information sub-working condition Cut.

[0049] Process parameter sub-working condition P:

[0050] During the machining process, different process parameters will cause changes in the cutting force, which will in turn affect the real-time health state of the CNC tool and have an impact on the prediction of the remaining service life. The present invention mainly summarizes the process parameter sub-working condition as cutting depth, spindle speed, feed rate, and feed mode: expressed as P = [ap n f way], where a p represents the cutting depth, n represents the spindle speed, f is the feed rate, and way represents the feed mode; the feed mode way is represented by one-hot encoding.

[0051] Sub-condition W of workpiece information:

[0052] The workpiece to be machined is the object in close contact with the NC tool. The workpiece information is the key factor affecting the stability of the machining process and also an important factor affecting the change of the NC tool state. The present invention formalizes the main parameters such as thermal conductivity, Young's modulus, friction coefficient, Poisson's ratio, tensile strength, shear strength, Rockwell hardness, impact toughness, melting point, etc. as consideration factors: expressed as W = [K E μs τ Rm ζ HRA Ak R], where K represents the thermal conductivity of the material, E represents the Young's modulus of the material, μs represents the friction coefficient, τ represents Poisson's ratio, Rm represents the tensile strength, ζ represents the shear strength, HRA represents Rockwell hardness, Ak represents impact toughness, and R represents the melting point of the material.

[0053] Sub-condition O of machine tool information:

[0054] Generally, the self-state of the machine tool equipment will inevitably affect the machining process. When predicting the remaining service life of the NC tool, this article takes the machine tool information as a necessary consideration condition, but does not consider the slight influence caused by the degradation and wear of the machine tool. Specifically expressed as O = [T in T out Sys Axle], where T in represents the internal environment temperature of the machine tool, T out represents the external environment temperature of the machine tool, Sys represents the machining accuracy of the machine tool, and Axle represents the number of spindles of the machine tool.

[0055] Sub-condition F of cutting fluid properties:

[0056] During the NC machining process, there is a certain correlation between the cutting fluid condition, cutting heat, and cutting force, which in turn affects the self-state of the cutting tool. Specifically expressed as F = [λ PH T F v], where λ represents the conductivity of the cutting fluid, PH represents the pH value of the cutting fluid, T F represents the cutting fluid temperature, and v represents the injection speed of the cutting fluid.

[0057] Sub-condition Cut of tool information:

[0058] As a decisive factor in the remaining service life of CNC tools, the tool's own attributes are of great significance in the assessment of tool health status. When considering life prediction in this invention, the working condition information of the tool's own attributes is 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 in degrees, L represents the tool overhang, and A represents the tool type.

[0059] Step 2: With the rapid development of sensor technology, data acquisition in the actual machining process is more efficient. Since the data collected during the machining process has the characteristics of a large amount of redundancy, it is necessary to preprocess, extract features, screen, and fuse the original data to avoid information loss caused by too high a feature dimension in the input of the prediction model and randomly selected features, weaken the generalization of the prediction model, and make it difficult to accurately predict the remaining service life of CNC tools under different working conditions.

[0060] Therefore, preprocess the monitoring signal data used in Step 1; perform time-domain feature extraction and frequency-domain feature extraction on the preprocessed monitoring signal data; then perform feature correlation analysis on the extracted features and the remaining service life of the tool to obtain strongly correlated features whose correlation with the remaining service life of the tool meets the requirements; and then perform feature fusion and dimensionality reduction on the obtained strongly correlated features to obtain a feature set sensitive to tool life.

[0061] Collect the monitoring signal data in the actual machining process of CNC tools, which covers processes such as the starting idle stroke stage, workpiece contact stage, full cut-in stage, stable cutting stage, tool cut-out stage, and retraction idle stroke stage during the machining process. The data signal segments have a large capacity. Also, because the actual machining cutting process is complex, the obtained signal data has characteristics such as a large amount of data, low value density, diverse sources, and high redundancy. These drawbacks weaken the data quality to varying degrees and reduce its own reliability. The premise for accurate and reliable prediction of the RUL of CNC tools is the authenticity and integrity of the data. Inputting data that does not meet the standard into the model will cause the prediction result to deviate too much, seriously affecting the prediction effect. Therefore, it is necessary to perform preliminary preprocessing on the original data in the early stage, including: effective value interception, missing value processing, data standardization, and Kalman filter processing.

[0062] (1) Effective value interception

[0063] Collect the complete machining data signals of the CNC tool from a new tool to scrapping, including situations such as machine stop, idle feed, tool change, and island avoidance. Inevitably, there are stages where the tool is not in contact with the workpiece. The data obtained in this case has no essential significance for predicting the remaining service life of the CNC tool. Therefore, it is necessary to appropriately intercept the overall signal data to improve the processing efficiency. This invention uses the boundary points of the sudden change in cutting force data as the upper and lower limits of the interception range, and takes the effective data in the stable cutting stage for analysis.

[0064] (2) Missing value processing

[0065] When acquiring data in a real processing scenario, missing data in the acquired data is usually caused by some uncontrollable factors (such as the degradation of machine tools, sudden changes in the environment, and the sensitivity of acquisition systems, etc.). According to the internal coding mechanism of computers, missing values are generally encoded as spaces, NaN, or other placeholders. Missing values will cause problems such as low model accuracy and prediction program errors in the subsequent use of data. In this embodiment, hierarchical mean imputation is adopted to complete the missing values in the monitoring signals. Hierarchical mean imputation accurately stratifies variables according to the inherent attribute characteristics of the variables before imputation, ensuring that the characteristics of the data in each layer are similar, and using the mean value of the complete units in each layer as the imputation value for that layer. The processing effect of the hierarchical mean imputation method is obvious, the supplemented data is reasonable, and it will not deviate too much from the distribution of the overall data, having a serious impact on the subsequent calculation and analysis.

[0066] (3) Data standardization

[0067] The data obtained during the processing process generally has different orders of magnitude and dimensions. When the level differences under different working conditions are large, a too large numerical magnitude will cause a certain factor to occupy too large a weight in the model, weakening the weight of some key factors that are significant but have a small numerical magnitude in the prediction model, seriously affecting the final effect of the prediction model. Therefore, it is necessary to perform standardization processing on the original data to eliminate the influence of dimensions and orders of magnitude on the subsequent calculation and analysis to the greatest extent, thereby ensuring the convergence speed and prediction accuracy of the model. In this embodiment, the Z-score method is used for standardization processing, which unifies the representation of dimensions and orders of magnitude, and at the same time does not lose the inherent attributes of the original data, retaining valuable information.

[0068] (4) Kalman filter processing

[0069] The unstable factors in the actual production environment cause the monitoring signals obtained by the intelligent tool holder to contain a relatively large amount of random noise. The key to its preprocessing is to ensure that the key information of the data itself is mined. Therefore, in this embodiment, the Kalman filter is used to update and process the data collected on-site in real time.

[0070] Aiming at the problem that the sampling points of the preprocessed signal are dense and the frequency is too high, making it difficult to directly input into the subsequent prediction model, in this embodiment, time-domain feature extraction and frequency-domain feature extraction are performed on the preprocessed monitoring signal data:

[0071] The time-domain features extracted from the preprocessed monitoring signal data include:

[0072]

[0073]

[0074] x in the table i (i = 1, 2, …, n) represents the sampling point sequence of the original signal.

[0075] The extracted frequency-domain features include:

[0076]

[0077] where f i represents the frequency corresponding to the cutting force signal sequence x i ; P i represents the power spectrum corresponding to f i ; represents the mean value of f i .

[0078] A total of 7 types of time-domain features and 5 types of frequency-domain features are extracted, totaling 12 types of time-frequency domain features. The composition of the cutting force signal contains information in a total of 3 dimensions, namely torque (Mon-torque), bending moment in the X direction (Mon-bending(x)), and bending moment in the Y direction (Mon-bending(y)), resulting in a feature matrix of 12×3 = 36. For the subsequent prediction model of the remaining service life of CNC tools, too high an input feature dimension will affect the prediction performance of the model, so further screening and optimization are required.

[0079] The maximum information coefficient method MIC is used to perform feature correlation analysis between the extracted features and the remaining service life of the tool, and strong correlation features that meet the requirements for the correlation with the remaining service life of the tool are obtained:

[0080] Step1: Using the monitoring signal data feature set {F i} extracted by time-domain and frequency-domain analysis, a total of 36 features {F1, F2, …, F 36} and the remaining service life of the tool {life i} form a new matrix D;

[0081] Step2: Using MIC analysis, all data points in D are distributed on a two-dimensional plane, and the plane is divided into grids by a×b straight lines. Calculate the probability distribution of each data point falling into the defined grid, which is called the "mutual information" of F and life under this division scheme. The formula is as follows:

[0082]

[0083] Step 3: Change the meshing scheme to obtain different "mutual information". After normalizing it to [0, 1], compare the mutual information values obtained under different-dimensional meshes, and take the maximum value MAX (Mutual Information) to determine the maximum information coefficient between the feature set and the remaining tool life, as shown in the following formula:

[0084]

[0085] In the formula: B is the present capacity function and B(n) = n α ; n is the number of experimental samples; α is the factor coefficient affecting the universality of the MIC method, and α is taken as 0.6.

[0086] Perform redundancy and correlation analysis on the total of 36 extracted features. Among the extracted features, the higher the correlation with the life feature, the more it can reflect the change process of the remaining tool life. The results are as follows:

[0087]

[0088]

[0089] For comparison, the Pearson Product-Moment Correlation Coefficient method (PPMCC) is also used here to perform feature correlation analysis on the extracted features and the remaining tool life. The results are as follows:

[0090]

[0091] It can be seen that the PPMCC analysis method discovers that 9 features have a correlation r = Corr(F, Y) with the remaining tool life with an absolute value greater than 0.6, that is, there are 9 strongly correlated features; the MIC analysis method discovers that 14 features have a correlation MIC[F; life] with the remaining tool life with an absolute value greater than 0.6, that is, there are 14 strongly correlated features. Among them, 9 features are consistent with those discovered by the PPMCC analysis method, and the remaining 5 features do not show a linear correlation with the remaining tool life but contain relatively valuable non-linear correlation information. Therefore, if the PPMCC analysis method is used, it will inevitably cause a large loss of non-linear correlation information. The present invention uses the MIC analysis method to comprehensively discover valuable and effective feature information.

[0092] The collected monitoring information is an important basis for reflecting the health status of CNC tools and a key analysis factor for the remaining service life of the tools. Due to the gradual increase in the capacity of the collected information, the characteristics of the monitoring signals increase accordingly, resulting in redundant characteristic information. If some characteristics are directly randomly selected, the value of the monitoring information will be lost. Therefore, the present invention uses the kernel principal component analysis method to perform feature fusion and dimensionality reduction on the obtained strongly correlated features, obtaining a feature set sensitive to the tool life, ensuring the value of the original information to the greatest extent, more comprehensively representing the internal information of the data, and avoiding the weakening of the generalization of the prediction model caused by excessive feature numbers or randomly selected features.

[0093] Step 3: Establish a Bi-GRU network model. For the current working condition, using the current working condition factors uniformly characterized in Step 1, the feature set sensitive to the tool life obtained in Step 2, the current state of the tool, and the remaining service life label corresponding to this working condition as the model training samples, construct a regression learner to train the Bi-GRU network model, and obtain the prediction model f for the remaining service life of the CNC tool under this working condition.

[0094] Step 4: When the prediction model f fails to face a new working condition task, adopt the incremental learning joint training method with L2 regularization, introduce a weight decay norm into the original loss function, and fine-tune the original prediction model f to obtain a "super model f'" that can predict both new task data and original task data, realizing the prediction of the RUL of the CNC tool under different working conditions.

[0095] The present invention uses a Bi-GRU network with incremental learning ability to predict the remaining service life of tools under different working conditions and gradually updates the network parameters during the learning process. Compared with the prediction method of the remaining service life of tools using a static learning model in the prior art, it has stronger applicability for predicting the remaining service life of tools under different working conditions. The Bi-GRU network based on the L2 regularization joint training method does not affect the originally learned knowledge after learning new samples in new knowledge, and the prediction error significantly decreases. That is, after learning new data, the prediction accuracy of the network is increased to more than 94%.

[0096] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and purposes of the present invention.

Claims

1. A method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network, characterized in that: It includes the following steps: Step 1: Collect the monitoring signal data during the actual machining process of the CNC tool, uniformly characterize the current working condition factors, and record the remaining service life label life and the machining quality constraint conditions corresponding to this working condition; Step 2: Preprocess the monitoring signal data used in Step 1; extract time-domain features and frequency-domain features from the preprocessed monitoring signal data; then perform feature correlation analysis between the extracted features and the remaining service life of the tool to obtain strongly correlated features whose correlation with the remaining service life of the tool meets the requirements; afterwards, perform feature fusion and dimensionality reduction on the obtained strongly correlated features to obtain a feature set sensitive to the tool life; Step 3: Establish a Bi-GRU network model. For the current working condition, use the current working condition factors uniformly characterized in Step 1, the feature set sensitive to the tool life obtained in Step 2, the current state of the tool, and the remaining service life label corresponding to this working condition as the model training samples, and construct a regression learner to train the Bi-GRU network model to obtain the prediction model f of the remaining service life of the CNC tool under this working condition; Step 4: When the prediction model f fails to face a new working condition task, use the incremental learning joint training method with L2 regularization to fine-tune the original prediction model f to obtain a "super model f'" that can predict both new task data and original task data, and realize the prediction of the RUL of the CNC tool under different working conditions.

2. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 1, wherein: In Step 1, the monitoring signal data during the actual machining process of the CNC tool are torque, bending moment in the X direction, and bending moment in the Y direction.

3. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 1 or 2, characterized in that: In Step 1, the working condition factors include the process parameter sub-working condition P, the workpiece information sub-working condition W, the machine tool information sub-working condition O, the cutting fluid property sub-working condition F, and the tool information sub-working condition Cut.

4. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 3, wherein: The process of uniformly characterizing the working condition factors is as follows: The process parameter sub-condition P, expressed as P = [a p n f way], where a p represents the cutting depth, n represents the spindle speed, f is the feed rate, and way represents the feed mode; the feed mode way is represented by one-hot encoding; The workpiece information sub-working condition W is expressed as W = [K E μs τ Rm ζ HRA Ak R], where K represents the thermal conductivity of the material, E represents the Young's modulus of the material, μs represents the friction coefficient, 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, expressed as O = [T in T out Sys Axle], where T in represents the internal environmental temperature of the machine tool, T out represents the external environmental temperature of the machine tool, Sys represents the machining accuracy of the machine tool, and Axle represents the number of spindles of the machine tool; Cutting fluid property sub-condition F, expressed as F = [λ PH T F v], where λ represents the conductivity of the cutting fluid, PH represents the pH value of the cutting fluid, T F represents the cutting fluid temperature, and v represents the injection speed of the cutting fluid; The tool information sub-working condition Cut is 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, L represents the tool overhang, and A represents the tool type.

5. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 1, wherein: In Step 1, the remaining service life of the CNC tool is characterized by the machinable travel, and the machining quality constraint condition is characterized by the surface roughness Ra of the workpiece to be machined.

6. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 1, wherein: In Step 2, the process of preprocessing the monitoring signal data includes: effective value interception, missing value processing, data standardization, and Kalman filter processing.

7. The method for predicting the remaining service life of a cutting tool under different working conditions based on a Bi-GRU network according to claim 1 or 6, characterized in that: In Step 2, the time-domain features extracted from the preprocessed monitoring signal data include mean, standard deviation, skewness, kurtosis, peak factor, root mean square, and waveform factor; the frequency-domain features extracted include power spectrum mean, power spectrum root mean square, power spectrum peak factor, power spectrum stability ratio, and power spectrum improved equivalent bandwidth.

8. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 1, characterized in that: In step 2, the maximum information coefficient method is used to perform feature correlation analysis on the extracted features and the remaining tool life, and strong correlation features whose correlation with the remaining tool life meets the requirements are obtained.

9. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 1, characterized in that: In step 2, the kernel principal component analysis method is used to perform feature fusion and dimensionality reduction on the obtained strong correlation features, and a feature set sensitive to tool life is obtained.

10. The method for predicting the remaining service life of a tool under different working conditions based on a Bi-GRU network according to claim 1, characterized in that: In step 4, the incremental learning joint training method using L2 regularization means: the L2 regularization method is used to introduce a weight decay norm into the original loss function.

Images