Cutter wear detection method

文档序号:123863 发布日期:2021-10-22 浏览:29次 中文

阅读说明:本技术 一种刀具磨损检测方法 (Cutter wear detection method ) 是由 王鸿亮 黄鹤翔 刘璐 朱湘宇 李航宇 于 2020-04-17 设计创作,主要内容包括:本发明涉及数控技术,神经网络技术,具体涉及一种刀具磨损检测方法。首先,采集高速刀具铣削时在不同轴向的振动和声发射信号,并进行数据预处理;然后,采用改进的3-K-Means聚类算法聚类出刀具的三种磨损状态区间,并提出多选择多隐层神经网络结构对其进行特征学习,再使用Softmax进行分类;最后,采用随机梯度下降对整个深层网络进行参数微调,建立刀具磨损检测模型。实验结果表明,所提出的方法在刀具磨损检测上准确率高达95%。(The invention relates to a numerical control technology and a neural network technology, in particular to a cutter abrasion detection method. Firstly, acquiring vibration and acoustic emission signals in different axial directions during high-speed cutter milling, and performing data preprocessing; then, clustering three wear state intervals of the cutter by adopting an improved 3-K-Means clustering algorithm, providing a multi-selection multi-hidden-layer neural network structure to perform characteristic learning on the cutter, and classifying the cutter by using Softmax; and finally, fine-tuning parameters of the whole deep network by adopting random gradient descent, and establishing a tool wear detection model. The experimental result shows that the accuracy of the method on the cutter abrasion detection is as high as 95%.)

1. A cutter wear detection method is characterized by comprising the following steps:

collecting a signal sample when a cutter is milled, and carrying out wavelet packet transformation on the signal sample to extract characteristics to obtain a label-free sample;

classifying the unlabeled samples by using an improved 3-K-Means clustering algorithm, calibrating the classified samples to obtain samples with labels, inputting the samples into a neural network for training to obtain a neural network model;

and acquiring signals of the cutter during milling in real time, extracting characteristics through wavelet analysis and wavelet packet transformation to obtain label-free data, and predicting and outputting a cutter abrasion state result according to a neural network model.

2. A tool wear detection method according to claim 1, wherein the signal samples are vibration and acoustic emission signals of the tool during milling.

3. The tool wear detection method according to claim 1, characterized in that a Softmax classifier is used for classifying the trained sample data, and a gradient descent algorithm is used for determining an optimal multi-hidden-layer neural network structure.

4. The tool wear detection method of claim 1, wherein the extracting features by wavelet analysis and wavelet packet transform is:

after wavelet packet transformation is carried out on the signal sample, the time sequence in the signal sample is decomposed into an approximation coefficient and a detail coefficient, whether the signal sample has singularity or not is judged according to the approximation coefficient and the detail coefficient, and when the signal sample has singularity, singularity features in the signal sample are extracted to serve as a label-free sample.

5. The tool wear detection method according to claim 1, wherein the unlabeled samples are classified by using a modified 3-K-Means clustering algorithm, and the classified samples are labeled as follows:

1) defining three initial clustering centers, and sequentially and respectively representing initial wear, normal wear and rapid wear;

2) each sample point, namely: assigning the unlabeled sample to the nearest clustering center to form a cluster; if the number of the points in the 2 nd cluster exceeds 50% of the total number of the unlabeled samples, keeping 50% of the points of the total number of the unlabeled samples which are closest to the cluster center of the 2 nd cluster, enabling the rest points to be classified into the 1 st or 3 rd cluster according to the cluster centers of the other two clusters, and updating the cluster center of each cluster;

3) and repeating the step 2) until the cluster is not changed or the maximum iteration number is reached.

6. The tool wear detection method according to claim 1 or 5, wherein the label is used to indicate: any one of initial wear, normal wear, and rapid wear.

7. The tool wear detection method of claim 1, wherein the tool wear state results are initial wear, normal wear, and rapid wear.

Technical Field

The invention relates to a numerical control technology and a neural network technology, in particular to a cutter abrasion detection method.

Background

In the high-speed milling process, the abrasion condition of the cutter is extremely complex, the abrasion state of the cutter is difficult to detect by a traditional artificial model method, and how to solve the problem by a more effective method becomes a research hotspot cutter in the field of intelligent processing. In the high-speed milling process, the abrasion state of the cutter is difficult to detect, and the machining precision and the product quality are influenced when the cutter is seriously abraded.

The traditional method mainly relies on the cutting force coefficient to analyze the abrasion of the cutter, needs to install an additional force measuring device and interferes the processing, so the problems of inevitable human factors exist. Meanwhile, the existing shallow model is more subjected to analysis and judgment aiming at a certain type or a certain signal, so that the existing shallow model is extremely easy to fall into local optimization, and in order to solve the problems of weak learning capacity and weak generalization capacity of the shallow model, the invention provides a cutter wear detection method based on 3-KMMBS.

Disclosure of Invention

The invention aims to provide a method for detecting the abrasion state of a cutter, which has high efficiency, high accuracy and strong stability and overcomes the defect of detecting the abrasion of the cutter by a shallow model.

The technical scheme adopted by the invention for realizing the purpose is as follows:

a tool wear detection method comprises the following steps:

collecting a signal sample when a cutter is milled, and carrying out wavelet packet transformation on the signal sample to extract characteristics to obtain a label-free sample;

classifying the unlabeled samples by using an improved 3-K-Means clustering algorithm, calibrating the classified samples to obtain samples with labels, inputting the samples into a neural network for training to obtain a neural network model;

and acquiring signals of the cutter during milling in real time, extracting characteristics through wavelet analysis and wavelet packet transformation to obtain label-free data, and predicting and outputting a cutter abrasion state result according to a neural network model.

The signal samples are vibration and acoustic emission signals when the tool is milled.

And classifying the trained sample data by using a Softmax classifier, and determining an optimal multi-hidden-layer neural network structure by using a gradient descent algorithm.

The extraction of the characteristics through wavelet analysis and wavelet packet transformation is as follows:

after wavelet packet transformation is carried out on the signal sample, the time sequence in the signal sample is decomposed into an approximation coefficient and a detail coefficient, whether the signal sample has singularity or not is judged according to the approximation coefficient and the detail coefficient, and when the signal sample has singularity, singularity features in the signal sample are extracted to serve as a label-free sample.

The method utilizes an improved 3-K-Means clustering algorithm to classify the unlabeled samples, and the classified samples are marked as follows:

1) defining three initial clustering centers, and sequentially and respectively representing initial wear, normal wear and rapid wear;

2) each sample point, namely: assigning the unlabeled sample to the nearest clustering center to form a cluster; if the number of the points in the 2 nd cluster exceeds 50% of the total number of the unlabeled samples, keeping 50% of the points of the total number of the unlabeled samples which are closest to the cluster center of the 2 nd cluster, enabling the rest points to be classified into the 1 st or 3 rd cluster according to the cluster centers of the other two clusters, and updating the cluster center of each cluster;

3) and repeating the step 2) until the cluster is not changed or the maximum iteration number is reached.

The label is used for representing: any one of initial wear, normal wear, and rapid wear.

The tool wear state results are initial wear, normal wear, and rapid wear.

The invention has the following beneficial effects and advantages:

1. the present invention improves upon the conventional K-Means algorithm for classifying the wear state of the cutting tool.

2. The invention abandons the structure of a shallow model and utilizes a multi-selection multi-hidden layer neural network for feature learning.

3. In the data preprocessing of the invention, wavelet packet variation and cepstrum analysis are used to extract high-frequency characteristics.

4. By using the method, the invention can realize accurate prediction of the wear state of the cutter. Through experimental verification, compared with the traditional machine learning method and a shallow model, the method has stronger learning capability and higher prediction accuracy.

Drawings

FIG. 1 is a graph of tool wear;

FIG. 2 is a flow chart of the 3-K-Means algorithm;

FIG. 3 is a diagram of a multi-choice multi-hidden layer neural network architecture;

FIG. 4 is a 3-KMMBS workflow diagram;

FIG. 5 is a comparison graph before and after the vibration signal transformation;

FIG. 6 is a raw signal and cepstral analysis;

FIG. 7 is the average accuracy of the 3-KMMBS predicted tool wear state.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and examples.

The characteristics extracted by the deep neural model are more natural and hierarchical, complex information in a high-dimensional data structure can be found, and the defect that the shallow model needs to be manually participated in extracting the characteristics of the sample is overcome. Meanwhile, the problem that the shallow model is easy to fall into local optimum is avoided by utilizing the pre-training process. In combination with the advantages of deep learning and the characteristics of tool wear, the invention provides a tool wear detection method based on 3-KMMBS (wherein 3-K represents a 3-K-Means algorithm, MB represents a Multi-selection Multi-hierarchy BP neural network, and S represents a Softmax classifier). In order to realize the above parts, the technical scheme adopted by the invention is as follows: a cutter detection method comprises the following steps:

firstly, clustering three wear state intervals of the cutter by using an improved 3-K-Means algorithm, extracting vibration and acoustic emission signals and the like by using Fourier transform, wavelet packet transform and the like as input parameters, performing characteristic learning on the three wear state intervals by adopting a multi-selection multi-hidden-layer BP neural network and the like, establishing a cutter wear detection model, and continuously optimizing the model through parameter fine tuning; finally, the proposed tool wear detection method is verified through experiments.

The present invention will be described in detail with reference to the accompanying drawings.

As shown in fig. 4, the present invention comprises the steps of:

firstly, acquiring a signal sample, and performing wavelet packet transformation on the signal sample to extract characteristics;

as shown in fig. 5, first, after db6 wavelet packet transformation is performed on the vibration signal, the sequence is decomposed into Approximation Coefficients (Approximation Coefficients) and Detail Coefficients (Detail Coeffi-centers), thereby extracting the singular characteristics thereof. Secondly, the spectrum analysis can embody the condition of how the signal or the time sequence is distributed along with the frequency, is convenient for experiment extraction of feature information which is difficult to observe, and is less influenced by artificial factors. Wherein the power spectrum can describe the variation of the signal power with frequency, and the cepstrum is the result of performing inverse fourier transform on the logarithm of the power spectrum. In the experiment, wavelet analysis is carried out on tool data in different wear states, and characteristic data in high and low frequencies are analyzed (original acoustic emission signals and acoustic emission signal cepstrum analysis are shown in fig. 6, wherein AE-RMS is the root mean square value of the acoustic emission signals).

Secondly, performing label-free sample classification by using an improved 3-K-Means clustering algorithm, labeling all label-free sample data, inputting the labeled sample data into a multi-selection multi-hidden-layer neural network, and performing supervised training on the labeled sample data;

the wear phase of the tool can be generally divided into an initial wear phase, a normal wear phase and a rapid wear phase. According to the tool wear condition in actual milling, the wear states of the tool can be classified into 3 types, and different wear states are given different label values, namely: the initial wear was label 0, the normal wear was label 1, and the sharp wear was label 2.

Among the clustering algorithms, the K-Means algorithm is improved compared with the classic and efficient contemporary K-Means algorithm, so that the proportion of three classes which are aggregated is as close to 1:5:4 as possible, and the algorithm is called as 3-K-Means algorithm. 3-K-Means algorithm idea: firstly, defining an initial clustering center; then each point is assigned to the nearest clustering center to form a cluster; if the number of the points in the 2 nd cluster exceeds 50% of the total number, classifying the points into the 1 st or 3 rd clusters according to the distance, and recalculating the cluster center of each cluster; and repeating the previous step until the cluster is not changed or the maximum iteration number is reached.

The clustering ratio is controlled to be about 1:5:4, because the ratio is obtained after a large number of experimental comparisons, and the tool wear curve obtained by the experiment is shown in fig. 1. By observing the curve, the clustering proportion of 1:5:4 can be found to better accord with the wear curve of the cutter and be closer to the real wear condition of the cutter. It can also be seen that in the early stages of the milling process the tool quickly reaches a state of light wear and the curve levels off relatively quickly near the boundary of light wear and moderate wear and that at the boundary of moderate wear and severe wear the curve suddenly becomes steeper and the slope increases relatively quickly.

As shown in FIG. 2, in 3-K-Means, the objective function is used to recalculate and how to compute the cluster center for each cluster, so the distance metric and objective function are the core of 3-K-Means. Here, Sum of Squared Errors (SSE) may be used as an objective function for clustering, as shown in equation (1).

Where K denotes the number of cluster centers, where K is 3. x denotes the sample coordinate to be calculated, ciDenotes the ith cluster center, dist denotes the euclidean distance between two points. For the k clustering center ckSolving, i.e. deriving SSE and making the derivative zero to obtain ckAs shown in formulas (2) and (3):

wherein m iskThe number of labeled samples for the kth class, namely: number of labeled samples of class k.

BP neural network, namely: the back propagation neural network is the most basic and the most representative neural network in deep learning. It is worth noting that in the cutter wear prediction process, the number of hidden layers and the number of nodes of the BP neural network limit the accuracy rate, so that a multi-choice and multi-hidden-layer strategy is provided for the BP neural network. Firstly, the hidden layer structure of each BP neural network is different, namely: the number of hidden layers, the number of nodes in each layer and the like are unique; secondly, vertically stacking a plurality of BP neural networks with different hidden layer structures for training; and finally, determining the optimal hidden layer structure by combining a Softmax classifier. By the method for adjusting and optimizing the hidden layer structure, the abrasion of the cutter can be accurately predicted.

Firstly inputting x1, x2, … and xmInputting into BP neural networks with different hidden layer structures (BP)1,BP2,……,BPn-1,BPn) And classifying by using a Softmax classifier, determining a deep network structure with the highest accuracy, and simultaneously using the multi-hidden-layer neural network for predicting the wear state of the tool in the future.

The whole working flow of the multi-selection multi-hidden-layer neural network is shown in figure 3.

As shown in fig. 7, in the third step, using the test sample set to classify the trained data by using a Softmax classifier, determining a multi-hidden-layer neural network with the highest accuracy, and using a gradient descent algorithm to perform parameter fine tuning and precision testing on the tool wear state prediction model.

The multi-hidden-layer neural network has strong feature learning capability as one of the structures of the deep neural network, but the classification prediction capability is not strong, and a classifier with the classification prediction capability needs to be combined. Among many classifiers, Softmax classifier has good classification prediction capability, and therefore is added to a multi-selection multi-hidden-layer neural network for classification prediction of tool wear state. The Softmax regression can solve the problem of multi-classification, and the model is a popularization of the logistic regression model on classification problems. Let the training set be { (x)i,yi) (ii) a 1,2, …, n, y ∈ {1,2, …, k }, assuming that the function is as shown in equation (4):

wherein, theta12,…,θkFor the model parameters to be trained, k is the number of classes, and for a given test input x, a probability value p (y j | x; θ) can be estimated for each class j using a hypothesis function.

The cost function of the Softmax regression model is shown in equation (5):

where m is the number of labeled samples.

11页详细技术资料下载
上一篇:一种医用注射器针头装配设备
下一篇:一种立铣刀在机磨损检测装置

网友询问留言

已有0条留言

还没有人留言评论。精彩留言会获得点赞!

精彩留言,会给你点赞!