CN111639461B - Tool wear state detection method aiming at industrial unbalanced data - Google Patents

Tool wear state detection method aiming at industrial unbalanced data Download PDF

Info

Publication number
CN111639461B
CN111639461B CN202010455079.2A CN202010455079A CN111639461B CN 111639461 B CN111639461 B CN 111639461B CN 202010455079 A CN202010455079 A CN 202010455079A CN 111639461 B CN111639461 B CN 111639461B
Authority
CN
China
Prior art keywords
data
training data
cutter
wear state
tool
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.)
Active
Application number
CN202010455079.2A
Other languages
Chinese (zh)
Other versions
CN111639461A (en
Inventor
刘振宇
刘惠
张朔
郏维强
谭建荣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202010455079.2A priority Critical patent/CN111639461B/en
Publication of CN111639461A publication Critical patent/CN111639461A/en
Application granted granted Critical
Publication of CN111639461B publication Critical patent/CN111639461B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/18—Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form
    • G05B19/406—Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form characterised by monitoring or safety
    • G05B19/4065—Monitoring tool breakage, life or condition
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/45—Nc applications
    • G05B2219/45136—Turning, lathe

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Human Computer Interaction (AREA)
  • Manufacturing & Machinery (AREA)
  • Automation & Control Theory (AREA)
  • Numerical Control (AREA)

Abstract

The invention discloses a cutter wear state detection method aiming at industrial unbalanced data. Preprocessing the historical monitoring data of the numerical control machine tool cutter obtained by the sensor and forming source training data by a cutter abrasion state label; the method comprises the following steps of performing same pretreatment on historical cutter monitoring data of a cutter to be detected and performing oversampling synthesis on source training data to obtain a few types of data to form auxiliary training data; training a tool wear state prediction model by using the training data through a transfer learning method; and preprocessing the real-time sensor data of the tool to be detected and inputting the preprocessed real-time sensor data into a prediction model to obtain the wear state of the tool in real time. The invention fully ensures the balance of the training data and the distribution consistency of the training data and the test data, thereby solving the problem of high-precision detection of the tool wear state under the conditions of less industrial data volume and imbalance.

Description

Tool wear state detection method aiming at industrial unbalanced data
Technical Field
The invention relates to a workpiece wear detection method, belonging to the field of numerical control machine tool state detection, in particular to a tool wear state detection method aiming at industrial unbalanced data, relating to the field of deep learning and migration learning unbalanced classification in machine learning.
Background
In modern machining and production operations, the wear of the tools of numerically controlled machine tools is a very common phenomenon. And the abrasion of the cutter directly influences the dimensional accuracy, the roughness and the quality of the processed surface, and even possibly leads to the scrapping of the processed workpiece, thereby increasing the production cost and reducing the processing efficiency. If the wear state of the numerical control machine tool can be monitored in real time, and the tool in the fault state can be replaced and maintained timely, the quality of a processed product can be effectively improved, the production efficiency is improved, the processing cost is reduced, and great economic benefits are achieved. In the present day that manufacturing systems tend to be flexible, online real-time monitoring of tool wear status is an important issue and is significant.
At present, the wear state of a numerical control machine tool is monitored mainly by adopting an indirect measurement method, namely, the wear state of the tool is predicted indirectly by processing a sensor signal in the machining process of the numerical control machine tool instead of a direct measurement mode, namely, whether the tool is in a normal state or a fault state is judged, and the method belongs to the two classification problems in machine learning. However, in actual industrial production, since the cost of collecting high-quality industrial data is high, the effective data amount that can be obtained is very small. In addition, in very limited data, the data distribution is also very unbalanced, i.e., the sample size in the fault state is much smaller than that in the normal state. This is because in industrial production, the number of devices operating in a fault state is much smaller than the number of devices operating in a normal state; the time length of the same device working in the fault state is far shorter than that in the normal state, so that the sample size of the fault state only accounts for 10 to 20 percent of that of the normal state, and even lower. Under the condition of extremely unbalanced data distribution, the traditional machine learning classification method is adopted, so that the classification result is inclined to a plurality of classes, namely a normal state. This is because the amount of failure data is too small, and the overall accuracy can be high even if all samples are classified as normal. Therefore, solving the problem of unbalanced distribution of industrial data is the central focus of solving the problem of monitoring the wear state of the tool.
And a few kinds of samples, namely fault state samples, are supplemented to serve as auxiliary training data, so that the idea of solving the problem of unbalanced distribution of industrial data is provided. On one hand, a few types of samples in the industrial data with similar cutter types and working conditions to the target task can be used as supplements; on the other hand, a few classes of samples may be generated as a complement. Such as the Synthesis of Minority Oversampling (SMOTE), is an effective method for synthesizing Minority samples from unbalanced data.
The problem of industrial data imbalance can be effectively solved by the two minority class data supplementing methods, namely adding the minority class samples in the similar industrial data and synthesizing the minority class samples by adopting the SMOTE algorithm, but a new problem is brought at the same time, namely the newly supplemented minority class data are inconsistent with the original data distribution in the target task, especially inconsistent with the test data distribution, and the reason is obvious. This poses a barrier to solving this problem with traditional machine learning algorithms, which can only give reasonable results under the condition that the training data and the test data are independently and identically distributed.
Transfer learning is an important means for solving the problem of distribution difference in the field of machine learning. It performs the migration of knowledge by reducing the distribution difference between the auxiliary domain to the target domain. Common migration learning methods include instance-based migration, feature-based migration, model-based migration, and relationship-based migration, among others. The example-based migration learning method allocates different weights to different samples according to the similarity degree of the auxiliary domain data samples and the target domain data samples to realize sample migration, and can effectively solve the problem that the supplemented few types of samples are not distributed consistently with the original data samples of the target task.
The neural network can effectively extract data characteristics, and has good effects in the fields of image recognition, natural language processing and the like. The gated cyclic unit network (GRU) can effectively capture the long-term dependence of each piece of data in sequence data on the data in the front and back directions, and has strong capability of processing the sequence data. The industrial sensor data is just sequence data, so the neural network, especially the gated cyclic unit network is very suitable for being used as a base learning device of the classification problem, namely a base classifier.
In the current research of applying a machine learning method to predict the wear state of a numerical control machine tool, the traditional machine learning method is mainly used. Such a method does not take into account the problems of serious shortage of high-quality sensor data and serious distribution imbalance in actual industrial production, and the obtained classification results tend to be in most categories. Therefore, the invention provides a tool wear state detection method aiming at industrial unbalanced data, which can effectively solve the problem of high-precision detection of the tool wear state under the conditions of small industrial data amount and unbalance.
Disclosure of Invention
The invention provides a tool wear state detection method aiming at industrial unbalanced data. The method fully ensures the balance of the training data by using the SMOTE algorithm, fully ensures the distribution consistency of the training data and the test data by using the transfer learning method, effectively improves the real-time detection effect of the wear state of the numerical control machine tool, and can be applied to the detection of the wear state of the numerical control machine tool in industrial production.
In order to realize the functions, the technical scheme of the invention specifically comprises the following technical steps:
s1, preprocessing the historical monitoring data of the numerical control machine tool cutter obtained through a sensor, and forming source training data with a corresponding cutter wear state label;
s2, carrying out the same pretreatment as that in the step S1 on the historical monitoring data of the cutter which is the same as the type of the cutter to be detected but different from the type of the cutter to be detected, wherein the historical monitoring data of the cutter and the corresponding cutter wear state label form part of auxiliary training data;
s3, oversampling is carried out on minority class data in the source training data by using an SMOTE algorithm, new minority class data are synthesized by using the oversampled data, and the new minority class data and a corresponding cutter wear state label form the other part of the auxiliary training data;
s4, forming training data by the source training data and the auxiliary training data, and training a tool wear state prediction model by using the training data through a transfer learning method;
and S5, preprocessing the real-time sensor data of the tool to be detected and inputting the preprocessed real-time sensor data into a prediction model to obtain the wear state of the tool in real time.
In the step S1, the historical cutter monitoring data comprises cutting force signals, vibration signals, sound signals and cutter abrasion loss in the cutter working process; the tool wear state includes a normal state and a fault state. The sensors include load cells, acceleration sensors and acoustic sensors.
The tool wear state is determined by setting a wear threshold according to the working conditions, data with the wear amount not less than the wear threshold is regarded as a fault state, and data with the wear amount less than the wear threshold is regarded as a normal state.
In the step S3, a few types of data in the tool history monitoring data refer to tool history sensor data in a fault state; the kind of tool history sensor data, the preprocessing method, and the generation method of the tool wear state flag in this step are the same as those in step S1.
In step S3, the SMOTE algorithm refers to a Synthetic minimal Oversampling Technique (Synthetic minimal Oversampling Technique), which analyzes a few classes of samples and synthesizes a new sample according to the few classes of samples to add to the data set, where the few classes of data and the tool wear status label refer to data of a fault state and the fault status label.
Since the sample size of the normal state and the sample size of the fault state in the source training data obtained by the sensor are greatly different, the sample size of the normal state is large, and the sample size of the fault state is small. After the processing in steps S2 and S3, the training data composed of the source training data in step S1 and the auxiliary training data in steps S2 and S3 can be equalized data, which means that the sample size in the normal state and the sample size in the fault state in the training data are not much different.
In step S4, the transfer learning method is described as follows:
s41, inputting: combining the auxiliary training data and the source training data to form a training data set T = T a ∪T b Number of iterations N, where T a N training data samples are used as auxiliary training data; t is b The source training data is m training data samples; the value range of the label of the tool wear state in the training data set is {0,1}, the label of 0 represents a normal state, and the label of 1 represents a fault state;
a group of historical cutter monitoring data and cutter wear state labels thereof form a training data sample;
s42, initializing parameters:
s421, establishing weight vectors of all training data samples
Figure BDA0002508933490000041
And initializing:
Figure BDA0002508933490000042
wherein the content of the first and second substances,
Figure BDA0002508933490000043
weights for the ith training data sample for the 1 st iteration; i.e. the weights of the n secondary training data samples are initialized to 1/n, and the weights of the m source training data samples are initialized to 1/m.
S422, initializing a weight update factor beta of the auxiliary training data:
Figure BDA0002508933490000044
s43, continuously iterating according to the following substeps, wherein the iteration is performed in N rounds, and in the t-th round of iteration, t =1, \8230, N:
s431, calculating the normalized weight distribution on the training data sample:
Figure BDA0002508933490000045
wherein w t Is the weight vector of the training data samples in the t-th iteration,
Figure BDA0002508933490000046
the weight of the ith training data sample in the t round of iteration;
s432, calling a base classifier, and distributing the weight p according to the training data set T and the current T-th iteration weight of the training data set T t To obtain a classifier h t ；
S433, calculating a classifier h t On-source training data T b Error rate of (2):
Figure BDA0002508933490000047
wherein x is i For the ith training data sample, h t (x i ) For input data samples of x i Predicted result of time, c (x) i ) For input data samples of x i Real result of time, h t (x i ) And c (x) i ) Both the value ranges are {0,1};
s434, calculating classifier weight:
β t ＝∈ t /(1-∈ t )
s435, updating the weight of the training data sample in the auxiliary training data:
Figure BDA0002508933490000048
s44, outputting a final classifier:
Figure BDA0002508933490000049
wherein x is the data to be tested.
In the step S4, the base classifier is a deep neural network composed of a bidirectional gated cyclic unit network and a fully connected network, and the deep neural network mainly comprises a layer of input layer, two layers of bidirectional gated cyclic unit networks, two layers of fully connected networks and an output layer in sequence; the loss function of deep neural network training is a cross entropy function, and an Adam optimization algorithm is adopted by an optimizer.
In step S5, the preprocessing method for the real-time sensor data is the same as the preprocessing method for the training data.
Under the conditions of limited industrial data scale and extremely unbalanced distribution, the method for supplementing minority samples from similar data and synthesizing the minority samples by adopting an SMOTE algorithm is adopted to supplement the original training data samples, so that the number of the minority samples is increased, and the problem of unbalanced industrial data is effectively solved; by adopting a transfer learning method, different weights are automatically distributed to training data samples according to the similarity between the auxiliary training data samples and the source training data samples, so that the problem that new training data and test data are not distributed consistently after a few types of samples are supplemented is effectively solved; and then the strong characteristic extraction and data analysis capabilities of the deep neural network are utilized, the high-precision detection of the abrasion state of the numerical control machine tool cutter is realized, and the method can be applied to the health management and equipment maintenance of the numerical control machine tool cutter.
Compared with the prior art and method, the method has the following advantages:
the invention effectively solves the two classification problems under the conditions of unbalanced data and inconsistent distribution of training data and test data. In actual industrial production, the high-quality industrial data quantity is seriously insufficient, and the fault data sample quantity is far smaller than the normal data sample quantity.
In the field of numerical control machine tool wear state prediction, the invention firstly proposes that original training data samples are supplemented by a method of supplementing a few samples from similar data and synthesizing the few samples by adopting an SMOTE algorithm, so that the number of the few samples is increased; and a transfer learning method is adopted, different weights are automatically distributed to training data samples according to the similarity between the auxiliary training data samples and the source training data samples, the problem that new training data and test data are not distributed consistently after a few types of samples are supplemented is solved, and finally high-precision prediction of the wear state of the numerical control machine tool is achieved.
The method provided by the invention fully ensures the balance of the training data, fully ensures the distribution consistency of the training data and the test data, solves the high-precision detection problem of the tool wear state under the condition of small and unbalanced industrial data quantity, and has better innovation and practicability.
Drawings
FIG. 1 is a schematic flow chart of the specific steps of the present invention.
Fig. 2 is a flow chart of the transfer learning method according to the present invention.
FIG. 3 is a schematic diagram of a neural network based classifier according to the present invention.
FIG. 4 is a graph showing the precision of the experimental results of the examples of the present invention.
FIG. 5 is a chart of the recall of the experimental results in the examples of the present invention.
FIG. 6 is a graph showing F1 values of experimental results in examples of the present invention.
Detailed Description
The invention is further illustrated with reference to the accompanying drawings and specific examples:
the specific embodiment of the invention is as follows:
the example uses the competition data of the 2010 PHM (fault diagnosis and health management) association data competition to verify the method for detecting the wear state of the cutter of the numerical control machine tool.
The processing parameters of the numerical control machine tool cutter are as follows: the rotating speed of the main shaft is 10400rpm, the feeding rate in the x-axis direction is 1555mm/min, the radial cutting depth is 0.125mm, and the axial cutting depth is 0.2mm. The sampling frequency of the measurement system was 50kHz. Each process consisted of 315 milling operations, which were stopped after each milling operation was completed and the amount of wear of the tool was measured using a LEICA MZ12 microscope. The data used in this example comprises data for 6 processes in total, resulting in 6 sets of sensor measurements, C1, C2, C3, C4, C5, C6 respectively.
The experimental task of the embodiment is to predict the wear state of the unmarked test data by utilizing the preprocessed tool sensor measurement data and the tool wear state label corresponding to the preprocessed tool sensor measurement data, compare the predicted result with the real wear state and quantitatively evaluate the result.
Fig. 1 shows a flow chart of the tool wear state detection method provided by the invention.
The specific implementation mode is as follows:
s1, setting training data and testing data:
in order to meet the basic requirements of the tool wear state detection method provided by the invention, match data are divided into a test data set, a source training data set and an auxiliary training data set. Where the source training dataset and the test dataset must be distributed consistently and the auxiliary training dataset is distributed similarly but not consistently to the test dataset.
Based on this requirement, three data sets C1, C4, C6 were selected as training and testing data for validating the proposed method of the present invention in the competition data. The specific method comprises the following steps: randomly extracting one third of data in one data set to serve as a test set; taking the remaining two thirds of data in the data set as a source training set; and taking a few class samples in the other two data sets as an auxiliary training set. In this way, since the test set and the source training set are from the same data set, it can be ensured that they are distributed consistently; and the auxiliary training set is derived from the other two data sets, so that the two data sets are similar to the distribution of the test set but inconsistent with the distribution of the test set.
Performing verification by adopting a three-fold verification mode, namely taking one data set as a test set and a source training set, and taking a few samples in the two data sets as an auxiliary training set; then the other data set is used as a test set and a source training set, a few class samples of the other two data sets are used as an auxiliary training set, and the like. The three sets of experiments were designated as experiment 1, experiment 2, and experiment 3, respectively, as shown in table 1.
TABLE 1
Figure BDA0002508933490000071
S2, establishing a cutter wear state label
Since the wear amount of the tool is given in the play data set, it is necessary to set a threshold value for the wear amount to convert it into a wear state and to label it as tool sensor data for this classification problem.
The specific method comprises the following steps: setting a wear loss threshold to 150 micrometers according to the working condition of the cutter in the data set and the condition of the data set, namely if the wear loss corresponding to one data sample is not less than 150 micrometers, considering that the data sample corresponds to a fault state and setting a wear state label of the data sample to be 1; if a data sample corresponds to a wear amount less than 150 microns, the data sample is considered to correspond to a normal state and its wear state label is set to 0. Obviously, the problem is a two-class problem, with class space of {0,1}.
S3, synthesizing a minority class data sample
And synthesizing a few types of samples, namely fault state samples, from the source training data set by adopting a SMOTE algorithm, wherein the total number of the samples is 90, and adding the samples into the auxiliary training data set.
In the steps S1 and S3, a small number of types of samples are supplemented in two different ways as auxiliary training data, so that the training data, i.e., the sum of the source training data and the auxiliary training data, can reach an equilibrium state, i.e., the difference between the normal state data amount and the fault state data amount is not large. Table 2 shows the normal state sample size and the fault state sample size in the training data samples of three experiments in one experiment (since the source training data samples are randomly extracted, the number of normal samples and fault samples in the source training set fluctuates in each experiment, but the amplitude is not large):
TABLE 2
Figure BDA0002508933490000072
S4, establishing a tool wear state prediction model by adopting a transfer learning method and using the training data formed in the steps S1, S2 and S3 and the corresponding wear state labels, wherein the attached figure 2 shows a flow schematic diagram of the transfer learning method in the invention, and the details are as follows:
s41, inputting: combining the auxiliary training data and the source training data to form a training data set T = T a ∪T b The number of iterations N =20, where T a N training data samples are used as auxiliary training data; t is b The source training data is m training data samples; the value range of the label of the tool wear state in the training data set is {0,1}, the label of 0 represents a normal state, and the label of 1 represents a fault state;
s42, initializing parameters:
s421. InitializationTraining data sample weight vectors
Figure BDA0002508933490000081
Wherein the content of the first and second substances,
Figure BDA0002508933490000082
wherein the content of the first and second substances,
Figure BDA0002508933490000083
weights for the ith training data sample for the 1 st iteration; i.e. the weights of the n secondary training data samples are initialized to 1/n, and the weights of the m source training data samples are initialized to 1/m.
S422, initializing weight updating factors of the auxiliary training data:
Figure BDA0002508933490000084
s43. Perform a total of N =20 iterations, in the t-th iteration (t =1, \8230;, N):
s431, calculating the normalized weight on the training data sample:
Figure BDA0002508933490000085
s432, calling a base classifier, and distributing p according to the training data set T and the weight of the current T-th iteration of the training data set T t To obtain a classifier h t ；
S433, calculating a classifier h t On-source training data T b Error rate of (2):
Figure BDA0002508933490000086
s434, calculating classifier weight:
β t ＝∈ t /(1-∈ t )
s435, updating the weight of the training data sample in the auxiliary training data:
Figure BDA0002508933490000087
s44, outputting a final classifier:
Figure BDA0002508933490000088
s5, constructing a deep neural network for realizing classification problem
The transfer learning method used in step S4 is essentially a boosting method, in which the base classifier called in step S432 is a deep neural network. Fig. 3 shows a schematic structural diagram of the neural network based classifier of the present invention, which is specifically as follows:
s51, constructing a bidirectional gating circulation unit sub-network, wherein the sub-network has two layers, the number of neurons in each layer of network is 15, the activating function adopts a 'tanh' function, and the input of the sub-network is preprocessed tool sensor measurement data;
s52, constructing a fully-connected sub-network, wherein the sub-network has two layers, the number of neurons in each layer of network is 15, the activation function of the sub-network adopts a linear rectification unit (ReLU), and the input of the sub-network is the output of the bidirectional gating circulation unit sub-network in the step S51;
s53, an output layer is constructed, the number of the neurons is 1, the output of the output layer represents the probability that the input data corresponds to the fault state, the activation function of the output layer adopts a sigmoid function, and the input of the output layer is the output of the fully-connected sub-network in the step S52;
s54, relevant parameters of neural network structure and model training:
the initialization method of all weights in the deep neural network adopts a normal distribution initialization method. The loss function of the neural network training is a cross entropy function, the optimizer adopts an Adam optimization algorithm, the learning rate is set to be 0.001, and the number of training rounds is set to be 20.
And S6, inputting the test data into the trained prediction model to predict the wear state of the cutter in real time.
Inputting the test data sample into the prediction model to obtain the predicted cutter wear state, comparing the predicted cutter wear state with the real cutter wear state label value, and calculating Precision ratio (Precision) and Recall ratio (Recall) of the prediction result and a blending average value (F1) based on the Precision ratio and the Recall ratio so as to quantitatively evaluate the prediction effect. The results statistics are shown in fig. 4, 5 and 6, respectively.
To reduce the effect of random errors, each set of experiments was repeated ten times, and the mean and variance of the ten experiments were calculated, with the results shown in table 3:
TABLE 3
Figure BDA0002508933490000091
As can be seen from table 3, through multiple experiments, the precision ratio, the recall ratio and the F1 value of the result of predicting the wear state of the tool of the numerical control machine tool by using the method provided by the present invention are all high, which indicates that the prediction method provided by the present invention has particularly good prediction accuracy. And the variance of the evaluation function value is smaller, which shows that the prediction model provided by the invention has better stability. In summary, the method provided by the invention is effective in the field of predicting the wear state of the tool of the numerically-controlled machine tool.
This example uses the data set used in the 2010 PHM International data Competition. Aiming at the problem of industrial data imbalance, a few classes are supplemented by introducing similar data and a synthetic method; and the problem of inconsistent distribution of test data and training data caused by supplementing data outside the domain is solved by a transfer learning method. Experimental results show that the prediction method provided by the invention has better accuracy and stability.
The invention provides the method for detecting the wear state of the numerical control machine tool cutter, which combines the balance of unbalanced data realized by supplementing similar data and using an SMOTE algorithm with the same distribution of test data and training data realized by a transfer learning method for the first time, effectively solves the problem of indirect monitoring of the wear state of the cutter under the condition of less and unbalanced industrial data quantity, obtains good effect, can be applied to the health management and maintenance of the numerical control machine tool cutter, and has innovation and practicability.
The above example is an application example of the present invention to a data set used in the 2010 PHM international data tournament, but the present invention is not limited to the above example. Similar solutions proposed according to the principles and concepts of the present invention should be considered as the protection scope of the present patent.

Claims (5)

1. A tool wear state detection method aiming at industrial unbalanced data is characterized by comprising the following steps:
s1, preprocessing the historical monitoring data of the numerical control machine tool cutter obtained through a sensor, and forming source training data with a corresponding cutter wear state label;
s2, carrying out the same pretreatment as that in the step S1 on the historical monitoring data of the cutter which is the same as the type of the cutter to be detected but different from the type of the cutter to be detected, wherein the historical monitoring data of the cutter and the corresponding cutter wear state label form part of auxiliary training data;
s3, oversampling is carried out on minority class data in the source training data by using an SMOTE algorithm, new minority class data are synthesized by using the oversampled data, and the new minority class data and a cutter wear state label corresponding to the new minority class data form the other part of the auxiliary training data;
s4, forming training data by the source training data and the auxiliary training data, and training a tool wear state prediction model by using the training data through a transfer learning method;
in step S4, the transfer learning method is described as follows:
s41, inputting: combining the auxiliary training data and the source training data to form a training data set T = T a ∪T b Number of iterations N, where T a N training data samples are used as auxiliary training data; t is a unit of b The source training data is m training data samples; the value range of the cutter wear state label in the training data set is {0,1}, and the label isA label of 0 indicates a normal state, and a label of 1 indicates a fault state;
s42, initializing parameters:
s421, establishing weight vectors of all training data samples
Figure FDA0003758873830000011
And initializing:
Figure FDA0003758873830000012
wherein the content of the first and second substances,
Figure FDA0003758873830000013
weights for the ith training data sample for the 1 st iteration;
s422, initializing a weight update factor beta of the auxiliary training data:
Figure FDA0003758873830000014
s43, continuously iterating according to the following substeps, wherein the iteration is carried out for N rounds, and in the t round of iteration, t =1, \8230, N:
s431, calculating the normalized weight distribution on the training data sample:
Figure FDA0003758873830000015
wherein, w t Is the weight vector of the training data samples in the t-th iteration,
Figure FDA0003758873830000016
the weight of the ith training data sample in the t round of iteration;
s432, calling a base classifier, and distributing the weight p according to the training data set T and the current T-th iteration weight of the training data set T t To obtain a classifier h t ；
S433, calculating a classifier h t On-source training data T b Error rate of (2):
Figure FDA0003758873830000021
wherein x is i For the ith training data sample, h t (x i ) For input data samples of x i Predicted result of time, c (x) i ) For input data samples of x i Real results of the time;
s434, calculating classifier weight:
β t ＝∈ t /(1-∈ t )
s435, updating the weight of the training data sample in the auxiliary training data:
Figure FDA0003758873830000022
s44, outputting a final classifier:
Figure FDA0003758873830000023
wherein x is the data to be tested;
and S5, preprocessing the real-time sensor data of the tool to be detected and inputting the preprocessed real-time sensor data into a prediction model to obtain the wear state of the tool in real time.
2. The tool wear state detection method for industrial imbalance data according to claim 1, wherein: in the step S1, the historical cutter monitoring data comprises cutting force signals, vibration signals, sound signals and cutter abrasion loss in the working process of the cutter; the tool wear state includes a normal state and a fault state.
3. The tool wear state detection method for industrial imbalance data according to claim 1, characterized in that: in step S3, the SMOTE algorithm refers to a Synthetic minimal Oversampling Technique (Synthetic minimal Oversampling Technique).
4. The tool wear state detection method for industrial imbalance data according to claim 1, wherein: in the step S4, the base classifier is a deep neural network composed of a bidirectional gated cyclic unit network and a fully connected network, and the deep neural network mainly comprises a layer of input layer, two layers of bidirectional gated cyclic unit networks, two layers of fully connected networks and an output layer in sequence; the loss function of deep neural network training is a cross entropy function, and an Adam optimization algorithm is adopted by an optimizer.
5. The tool wear state detection method for industrial imbalance data according to claim 1, characterized in that: in step S5, the preprocessing method for the real-time sensor data is the same as the preprocessing method for the training data.
CN202010455079.2A 2020-05-26 2020-05-26 Tool wear state detection method aiming at industrial unbalanced data Active CN111639461B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010455079.2A CN111639461B (en) 2020-05-26 2020-05-26 Tool wear state detection method aiming at industrial unbalanced data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010455079.2A CN111639461B (en) 2020-05-26 2020-05-26 Tool wear state detection method aiming at industrial unbalanced data

Publications (2)

Publication Number Publication Date
CN111639461A CN111639461A (en) 2020-09-08
CN111639461B true CN111639461B (en) 2022-10-18

Family

ID=72331147

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010455079.2A Active CN111639461B (en) 2020-05-26 2020-05-26 Tool wear state detection method aiming at industrial unbalanced data

Country Status (1)

Country Link
CN (1) CN111639461B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112192318B (en) * 2020-09-28 2022-07-19 北京天泽智云科技有限公司 Machining tool state monitoring method and system
CN112257767B (en) * 2020-10-16 2023-03-17 浙江大学 Product key part state classification method aiming at class imbalance data
CN112518425B (en) * 2020-12-10 2022-10-04 南京航空航天大学 Intelligent machining cutter wear prediction method based on multi-source sample migration reinforcement learning
CN112801139B (en) * 2021-01-06 2024-04-26 南京航空航天大学 Intelligent cutter wear state identification method based on heterogeneous domain self-adaptive transfer learning
CN113084237B (en) * 2021-04-16 2022-07-29 贵州大学 Method and device for predicting wear value of milling cutter, electronic device and storage medium
CN115319538B (en) * 2022-08-30 2023-11-17 上海诺倬力机电科技有限公司 Five-shaft head cutter abrasion intelligent detection method and system

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5969676B1 (en) * 2015-09-30 2016-08-17 ファナック株式会社 Machine learning device and machine learning method for optimizing frequency of tool correction of machine tool, and machine tool including the machine learning device
WO2017192821A1 (en) * 2016-05-06 2017-11-09 Massachusetts Institute Of Technology Method and apparatus for efficient use of cnc machine shaping tool including cessation of use no later than the onset of tool deterioration by monitoring audible sound during shaping
US11531852B2 (en) * 2016-11-28 2022-12-20 D-Wave Systems Inc. Machine learning systems and methods for training with noisy labels
JP6487475B2 (en) * 2017-02-24 2019-03-20 ファナック株式会社 Tool state estimation device and machine tool
CN107644057B (en) * 2017-08-09 2020-03-03 天津大学 Absolute imbalance text classification method based on transfer learning
CN109822399B (en) * 2019-04-08 2020-07-14 浙江大学 Numerical control machine tool cutter wear state prediction method based on parallel deep neural network
CN111037365B (en) * 2019-12-26 2021-08-20 大连理工大学 Cutter state monitoring data set enhancing method based on generative countermeasure network

Also Published As

Publication number Publication date
CN111639461A (en) 2020-09-08

Similar Documents

Publication Publication Date Title
CN111639461B (en) Tool wear state detection method aiming at industrial unbalanced data
Chen et al. Predicting tool wear with multi-sensor data using deep belief networks
CN110355608B (en) Cutter wear amount prediction method based on self-attention mechanism and deep learning
CN110153802B (en) Tool wear state identification method based on convolution neural network and long-term and short-term memory neural network combined model
Guo et al. Online remaining useful life prediction of milling cutters based on multisource data and feature learning
CN112883994B (en) Rotating machinery variable working condition fault diagnosis method with balanced distribution adaptation
CN113255848B (en) Water turbine cavitation sound signal identification method based on big data learning
CN112257530B (en) Rolling bearing fault diagnosis method based on blind signal separation and support vector machine
CN110561191B (en) Numerical control machine tool cutter abrasion data processing method based on PCA and self-encoder
CN112257767B (en) Product key part state classification method aiming at class imbalance data
Ayodeji et al. Causal augmented ConvNet: A temporal memory dilated convolution model for long-sequence time series prediction
CN114619292B (en) Milling cutter wear monitoring method based on fusion of wavelet denoising and attention mechanism with GRU network
CN108581633A (en) A method of based on the more sensor monitoring cutting tool states of genetic algorithm optimization
CN114676742A (en) Power grid abnormal electricity utilization detection method based on attention mechanism and residual error network
CN113188794B (en) Gearbox fault diagnosis method and device based on improved PSO-BP neural network
CN112611584A (en) Fatigue failure detection method, device, equipment and medium for wind generating set
Xie et al. Tool condition prognostic model based on digital twin system
CN114905335A (en) Cutter wear prediction method combining domain confrontation and convolution neural network
CN115112372A (en) Bearing fault diagnosis method and device, electronic equipment and storage medium
CN113103068B (en) Cutter state monitoring method based on deep migration learning
CN112763215B (en) Multi-working-condition online fault diagnosis method based on modular federal deep learning
CN113221946A (en) Method for diagnosing fault types of mechanical equipment
CN116372665A (en) OPC UA-based aviation numerical control machine tool health monitoring method
Wei et al. Research on tool wear state identification method driven by multi-source information fusion and multi-dimension attention mechanism
Elbestawi et al. Tool condition monitoring in machining

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
GR01 Patent grant
GR01 Patent grant