WO2021046738A1 - Method for monitoring state of deep hole boring cutter on basis of stacked autoencoder - Google Patents
Method for monitoring state of deep hole boring cutter on basis of stacked autoencoder Download PDFInfo
- Publication number
- WO2021046738A1 WO2021046738A1 PCT/CN2019/105283 CN2019105283W WO2021046738A1 WO 2021046738 A1 WO2021046738 A1 WO 2021046738A1 CN 2019105283 W CN2019105283 W CN 2019105283W WO 2021046738 A1 WO2021046738 A1 WO 2021046738A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- autoencoder
- layer
- data
- deep hole
- tool
- Prior art date
Links
Images
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
- B23Q—DETAILS, COMPONENTS, OR ACCESSORIES FOR MACHINE TOOLS, e.g. ARRANGEMENTS FOR COPYING OR CONTROLLING; MACHINE TOOLS IN GENERAL CHARACTERISED BY THE CONSTRUCTION OF PARTICULAR DETAILS OR COMPONENTS; COMBINATIONS OR ASSOCIATIONS OF METAL-WORKING MACHINES, NOT DIRECTED TO A PARTICULAR RESULT
- B23Q17/00—Arrangements for observing, indicating or measuring on machine tools
- B23Q17/09—Arrangements for observing, indicating or measuring on machine tools for indicating or measuring cutting pressure or for determining cutting-tool condition, e.g. cutting ability, load on tool
Definitions
- the invention belongs to the technical field of tool condition monitoring, and specifically relates to a method for monitoring the condition of a deep hole boring tool based on a stacked self-encoder.
- Deep holes In the machinery manufacturing industry, cylindrical holes with a hole depth of more than 5 times the hole diameter are generally called deep holes.
- Commonly used processing methods of deep holes mainly include drilling, reaming, boring, reaming and so on.
- the boring process is widely used for deep hole processing of larger structure size.
- the deep hole obtained by boring has high precision and low cost. Therefore, many manufacturing enterprises adopt the boring process.
- the cutting area of boring is located inside the deep hole, and the wear and damage state of the tool is difficult to observe with the naked eye.
- the machine tool operator can only rely on experience to determine the cutting state of the tool inside the deep hole by observing the outflow of chips and the touch of the hand to perceive the vibration of the boring bar. It is difficult to accurately determine the real-time state of the tool. When there are abnormal situations such as tool breakage, chipping, etc., failure to make timely judgments and take corresponding measures may lead to scrapped parts.
- a two-level deep learning model including a deep confidence network and a convolutional neural network is constructed, based on a large number of CNC machining monitoring signals Train the deep learning network to realize real-time monitoring of the state of CNC machining tools.
- the image data acquisition module is used to shoot the video of the tool cutting workpiece during the cutting process, and the image data preprocessing module is used to extract the video The extracted image is positioned, cropped and normalized.
- the edge calculation module integrated with the knife discriminator is used to receive the processed image, and the pre-trained convolutional neural network forward inference is used to obtain the knife Determine the result.
- the present invention proposes a real-time monitoring method of deep hole boring tool condition based on a stacked autoencoder.
- the purpose of the present invention is to provide a method for effectively monitoring the real-time status of the boring tool, and solve the problem that the status of the boring tool is difficult to monitor.
- the technical scheme of the present invention is as follows: firstly, two three-way acceleration sensors are respectively adsorbed on the outside of the two cage bushes of the deep hole boring bar through the magnetic base, and one is placed at the processing entrance of the inner hole of the workpiece.
- the microphone collects the vibration and sound signals during the boring process; then, uses the limiting filter method to preprocess the collected data; then, builds a stacked autoencoder network, adopts a greedy layer-by-layer method, and uses preprocessing
- the latter data trains the stacked autoencoder; finally, the real-time vibration and sound signals in the boring process are preprocessed into the stacked autoencoder network, and the network outputs the current state of the tool, thereby realizing deep hole boring Real-time monitoring of knife status.
- a method for monitoring the state of a deep hole boring tool based on a stacked autoencoder which is characterized in that the steps are as follows:
- the first step is to collect vibration and sound information during the deep hole boring process
- the two three-directional acceleration sensors are adsorbed on the two cage bearings of the deep hole boring bar through the magnetic base, and the microphone is placed at one end of the inner hole of the workpiece to collect the tool bar vibration and cutting noise during the machining process;
- the data collected by the three-way acceleration sensor is segmented according to the sampling frequency, and each segment of data is set to x(n), and fast Fourier transform is performed on each segment of data:
- k 0,...,N-1;
- the third step is the construction and training of the stacked autoencoder network
- f( ⁇ ) represents the activation function of the neuron, Represents the weight of the input layer and the hidden layer, Indicates the weight of the hidden layer and the output layer, Shows the bias of each neuron in the hidden layer, Represents the bias of each neuron in the output layer;
- I the output data of the autoencoder
- ⁇ is the sparse penalty term parameter, which is used to control the proportion of the sparse penalty term in the loss function
- ⁇ is the sparsity parameter
- the greedy training method is used for unsupervised training of the autoencoder, the training set is input to the first layer of autoencoder, and it is trained to minimize the loss function Loss to obtain the optimal weight and bias; the first layer of autoencoder
- the hidden layer output of the second-layer autoencoder is used as the input of the second-layer autoencoder to train it.
- the output of the second-layer autoencoder is used to perform supervised training on the softmax classifier, and then the trained two-layer autoencoder
- the encoder and the softmax classifier are stacked to obtain a trained stacked autoencoder; after training, the remaining test set is used to test the stacked autoencoder; when the test accuracy is higher than 90%, the model can be used for tool status monitor;
- the fourth step real-time monitoring of deep hole boring tool status
- the real-time data is input into the trained stacked autoencoder network model after data preprocessing, and the model outputs the real-time status of the tool; when the tool status is normal, the model output is 1; when the tool status is off When cutting, the model output is 2; when the tool is blunt, the model output is 3.
- the beneficial effects of the invention through the method, real-time monitoring of the state of the deep hole boring tool is realized, the dependence on the experience of the machine tool operator is reduced, the processing efficiency is improved, and the rejection rate is reduced.
- Figure 1 is a flow chart of deep hole boring tool status monitoring.
- Figure 2 is a schematic diagram of the sensor layout of a deep hole boring machine.
- Figure 3 is a structural diagram of the first-layer self-encoder.
- Figure 4 is a structural diagram of a stacked self-encoder.
- Figure 5 shows the broken tool monitoring waveform.
- Figure 6 is a blunt monitoring waveform diagram.
- the first step is to collect vibration and sound information during the deep hole boring process
- the #1 three-way acceleration sensor 5 and #2 three-way acceleration sensor 7 are attached to the side of the tool holder bearing bush with a magnetic seat, and the installation method is shown in Figure 2. Set the sampling frequency to 1000 Hz to collect vibration and sound signals during processing.
- the frequency domain data is classified according to three states: normal, broken, and blunt. Each sample contains 7000 data. Among them, the number of samples in the normal state is 20000, and the number of samples in the broken and worn states are 5000 and 3000 respectively.
- the third step is the construction and training of the stacked autoencoder network
- Stack two autoencoders and one softmax classifier to construct a stacked autoencoder network The number of neurons in the input layer and output layer of the first autoencoder is 7000, and the number of neurons in the hidden layer is 3000, as shown in Figure 3.
- the number of neurons in the input and output layers of the second autoencoder is 3000, and the number of neurons in the hidden layer is 1000.
- the number of neurons in the input and output layer of the third autoencoder is 1000, and the number of neurons in the hidden layer is 500. Randomly select 2000 samples in each of the three states to train the autoencoder, and use the greedy training method to train the autoencoder layer by layer.
- the fourth step real-time monitoring of deep hole boring tool status
- the real-time vibration and sound data are preprocessed and input into the stacked autoencoder model to monitor the tool status during the boring process. As shown in Figures 5 and 6, the model can accurately judge the real-time status of the tool.
Abstract
A method for monitoring the state of a deep hole boring cutter on the basis of a stacked autoencoder, the method comprising: first using magnetic bases to adhere two three-way acceleration sensors (5, 7) on two holder bearings of a deep hole boring rod (6), respectively, and placing a microphone (3) at a machining inlet of a workpiece (1) inner hole so as to collect vibration signals and sound signals during boring; using amplitude limiting filtering to perform data preprocessing on the collected data; then, constructing a stacked autoencoder network model, and using a greed layer-wise method to train a stacked autoencoder by using the preprocessed data; and finally, inputting real-time vibration signals and sound signals during boring machining into the stacked autoencoder network model after data preprocessing, the network model outputting the real-time state of a cutter. The described method may achieve the real-time monitoring of the state of a deep hole boring cutter.
Description
本发明属于刀具状态监测技术领域，具体为一种基于堆叠自编码器的深孔镗刀状态监测方法。The invention belongs to the technical field of tool condition monitoring, and specifically relates to a method for monitoring the condition of a deep hole boring tool based on a stacked self-encoder.
在机械制造业中，一般将孔深超过孔径5倍的圆柱孔称为深孔。常用的加工深孔的工艺方法主要有钻孔、扩孔、镗孔、铰孔等。目前，较大结构尺寸的深孔加工广泛采用的是镗削工艺，镗孔得到的深孔精度高、成本较低，因此很多的制造业企业都采用镗削工艺。与一般的车削、铣削相比，镗削的切削区域位于深孔内部，刀具的磨损破损状态很难用肉眼进行观察。机床操作者只能凭借经验，通过观察流出的切屑、手的触觉感知镗杆的振动，来确定深孔内部刀具的切削状态，难以准确判断刀具的实时状态。当出现刀具断裂、崩刃等异常情况时，不能及时做出判断并采取相应措施可能会导致零件的报废。In the machinery manufacturing industry, cylindrical holes with a hole depth of more than 5 times the hole diameter are generally called deep holes. Commonly used processing methods of deep holes mainly include drilling, reaming, boring, reaming and so on. At present, the boring process is widely used for deep hole processing of larger structure size. The deep hole obtained by boring has high precision and low cost. Therefore, many manufacturing enterprises adopt the boring process. Compared with general turning and milling, the cutting area of boring is located inside the deep hole, and the wear and damage state of the tool is difficult to observe with the naked eye. The machine tool operator can only rely on experience to determine the cutting state of the tool inside the deep hole by observing the outflow of chips and the touch of the hand to perceive the vibration of the boring bar. It is difficult to accurately determine the real-time state of the tool. When there are abnormal situations such as tool breakage, chipping, etc., failure to make timely judgments and take corresponding measures may lead to scrapped parts.
在专利“一种基于深度学习的复杂结构件数控加工刀具状态实时监测方法”(CN201710739173.9)中，构建了包含深度置信网络和卷积神经网络两级深度学习模型，基于大量数控加工监测信号训练深度学习网络，进而实现数控加工刀具状态的实时监测。在专利“一种基于深度学习的数控机床断刀检测系统及方法”(CN201910228970.X)中，利用图像数据采集模块拍摄切削加工过程中刀具切削工件的视频，利用图像数据预处理模块提取视频中的图像，并对提取的图像进行定位、裁剪和归一化处理，利用集成有断刀判别器的边缘计算模块接收处理后的图像，并利用预先训练的卷积神经网络前向推理得到断刀判别结果。In the patent "A method for real-time monitoring of tool status in CNC machining of complex structural parts based on deep learning" (CN201710739173.9), a two-level deep learning model including a deep confidence network and a convolutional neural network is constructed, based on a large number of CNC machining monitoring signals Train the deep learning network to realize real-time monitoring of the state of CNC machining tools. In the patent "A deep learning-based CNC machine tool break detection system and method" (CN201910228970.X), the image data acquisition module is used to shoot the video of the tool cutting workpiece during the cutting process, and the image data preprocessing module is used to extract the video The extracted image is positioned, cropped and normalized. The edge calculation module integrated with the knife discriminator is used to receive the processed image, and the pre-trained convolutional neural network forward inference is used to obtain the knife Determine the result.
通过分析上述专利可知，基于图像识别或切削力信号的刀具状态监测方法无法用于深孔镗削状态监测。本发明针对深孔镗削刀具状态监测的难题，提出 一种基于堆叠自编码器的深孔镗刀状态实时监测方法。By analyzing the above-mentioned patents, it can be known that the tool condition monitoring method based on image recognition or cutting force signal cannot be used for deep hole boring condition monitoring. Aiming at the difficult problem of deep hole boring tool condition monitoring, the present invention proposes a real-time monitoring method of deep hole boring tool condition based on a stacked autoencoder.
发明内容Summary of the invention
本发明的目的为提供一种有效监测镗刀实时状态的方法，解决镗刀状态难以监测的问题。The purpose of the present invention is to provide a method for effectively monitoring the real-time status of the boring tool, and solve the problem that the status of the boring tool is difficult to monitor.
为解决上述技术问题，本发明的技术方案为：首先，将两个三向加速度传感器通过磁座分别吸附在深孔镗杆的两个保持架轴瓦外部，在工件内孔的加工进口处放置一个传声器，采集镗削过程中的振动和声信号；然后，采用限幅值滤波法，对采集到的数据进行数据预处理；接着，构建堆叠自编码器网络，采用贪婪逐层方法，利用预处理后的数据对堆叠自编码器进行训练；最后，将镗削加工过程中的实时振动和声信号经数据预处理后输入堆叠自编码器网络中，网络输出刀具的当前状态，从而实现深孔镗刀状态的实时监测。In order to solve the above technical problems, the technical scheme of the present invention is as follows: firstly, two three-way acceleration sensors are respectively adsorbed on the outside of the two cage bushes of the deep hole boring bar through the magnetic base, and one is placed at the processing entrance of the inner hole of the workpiece. The microphone collects the vibration and sound signals during the boring process; then, uses the limiting filter method to preprocess the collected data; then, builds a stacked autoencoder network, adopts a greedy layer-by-layer method, and uses preprocessing The latter data trains the stacked autoencoder; finally, the real-time vibration and sound signals in the boring process are preprocessed into the stacked autoencoder network, and the network outputs the current state of the tool, thereby realizing deep hole boring Real-time monitoring of knife status.
一种基于堆叠自编码器的深孔镗刀状态监测方法，其特征在于，步骤如下：A method for monitoring the state of a deep hole boring tool based on a stacked autoencoder, which is characterized in that the steps are as follows:
第一步，深孔镗削过程中的振动和声音信息采集The first step is to collect vibration and sound information during the deep hole boring process
将两个三向加速度传感器通过磁座吸附在深孔镗杆的两个保持架轴承上，将传声器放置于工件内孔的一端，对加工过程中的刀杆振动以及切削噪声进行采集；The two three-directional acceleration sensors are adsorbed on the two cage bearings of the deep hole boring bar through the magnetic base, and the microphone is placed at one end of the inner hole of the workpiece to collect the tool bar vibration and cutting noise during the machining process;
第二步，振动和声音数据预处理The second step, vibration and sound data preprocessing
对三向加速度传感器采集到的数据按照采样频率进行分段，设每段数据为x(n)，对每段数据做快速傅里叶变换：The data collected by the three-way acceleration sensor is segmented according to the sampling frequency, and each segment of data is set to x(n), and fast Fourier transform is performed on each segment of data:
其中，k＝0,...,N-1；Among them, k=0,...,N-1;
计算其单侧幅频谱，将幅值低于0.2的数据全部滤掉；将滤波后的数据按照正常、磨钝和断刀三种状态进行分组；Calculate the one-sided amplitude spectrum, filter out all the data with amplitude less than 0.2; group the filtered data according to the three states of normal, blunt and broken;
第三步，堆叠式自编码器网络的构建和训练The third step is the construction and training of the stacked autoencoder network
取相同数量的三种状态下的样本数据做训练集，剩余样本做测试集，训练集和测试集要均包含三种状态下的样本数据；设输入数据为x＝[x
1,x
2,…,x
i]，则自编码器的输入输出关系为：
Take the same number of sample data in the three states as the training set, and the remaining samples as the test set. Both the training set and the test set must contain the sample data in the three states; set the input data as x=[x 1 ,x 2 , …,X i ], the input and output relationship of the autoencoder is:
式中，f(·)表示神经元的激活函数，
表示输入层与隐含层的权重，
表示隐含层与输出层的权重，
示隐含层各神经元的偏置，
表示输出层各神经元的偏置；
In the formula, f(·) represents the activation function of the neuron, Represents the weight of the input layer and the hidden layer, Indicates the weight of the hidden layer and the output layer, Shows the bias of each neuron in the hidden layer, Represents the bias of each neuron in the output layer;
构造损失函数：Construct the loss function:
式中，
是自编码器的输出数据；α是稀疏惩罚项参数，作用是控制稀疏惩罚项在损失函数当中所占的比重；γ是稀疏性参数；
是隐含层第j个神经元的激活度；
Where Is the output data of the autoencoder; α is the sparse penalty term parameter, which is used to control the proportion of the sparse penalty term in the loss function; γ is the sparsity parameter; Is the activation degree of the jth neuron in the hidden layer;
利用亚当优化器，最小化损失函数，优化权重w
ij和偏置b：
Using Adam optimizer, minimize the loss function, optimize the weight w ij and bias b:
式中，t为迭代次数；η为学习率；
和
分别为一阶动量项和二阶动量 项；λ
1和λ
2为动力值分别取0.9和0.999；
和
分别为一阶和二阶动量项的修正值；(w
ij,b)
t表示第t次迭代时的模型权重和偏置；g
t＝ΔJ((w
ij,b)
t)表示t次迭代时代价函数关于(w
ij,b)
t的梯度；∈为常数，避免分母为0，取值很小；
In the formula, t is the number of iterations; η is the learning rate; with They are the first-order momentum term and the second-order momentum term respectively; λ 1 and λ 2 are the dynamic values, taking 0.9 and 0.999 respectively; with Respectively are the correction values of the first-order and second-order momentum terms; (w ij ,b) t represents the model weight and bias at the tth iteration; g t =ΔJ((w ij ,b) t ) represents t iterations The gradient of the time cost function with respect to (w ij ,b) t ; ∈ is a constant, to avoid the denominator being 0, and the value is small;
采用贪婪训练法对自编码器进行无监督训练，将训练集输入第一层自编码器，对其进行训练，最小化损失函数Loss，得到最优的权重和偏置；将第一层自编码器的隐层输出作为第二层自编码器的输入，对其进行训练，训练完成后用第二层自编码器的输出对softmax分类器，进行有监督训练，再将训练好的两层自编码器和softmax分类器堆叠起来，得到训练好的堆叠自编码器；训练完毕后，采用剩余测试集对堆叠自编码器进行测试；当测试准确率高于90％时，模型可用于刀具状态的监测；The greedy training method is used for unsupervised training of the autoencoder, the training set is input to the first layer of autoencoder, and it is trained to minimize the loss function Loss to obtain the optimal weight and bias; the first layer of autoencoder The hidden layer output of the second-layer autoencoder is used as the input of the second-layer autoencoder to train it. After the training is completed, the output of the second-layer autoencoder is used to perform supervised training on the softmax classifier, and then the trained two-layer autoencoder The encoder and the softmax classifier are stacked to obtain a trained stacked autoencoder; after training, the remaining test set is used to test the stacked autoencoder; when the test accuracy is higher than 90%, the model can be used for tool status monitor;
第四步，深孔镗刀状态的实时监测The fourth step, real-time monitoring of deep hole boring tool status
在实际加工过程中，将实时数据经数据预处理后输入至训练好的堆叠自编码器网络模型中，模型输出刀具的实时状态；当刀具状态正常时，模型输出为1；当刀具状态为断刀时，模型输出为2；当刀具状态为磨钝时，模型输出为3。In the actual machining process, the real-time data is input into the trained stacked autoencoder network model after data preprocessing, and the model outputs the real-time status of the tool; when the tool status is normal, the model output is 1; when the tool status is off When cutting, the model output is 2; when the tool is blunt, the model output is 3.
本发明的有益效果：通过该方法，实现了深孔镗刀状态的实时监测，降低对机床操作工经验的依赖，提高加工效率，降低废品率。通过对原始数据的滤波去除冗余数据，加快堆叠自编码器的训练速度，提高预测的准确度；在损失函数中加入稀疏惩罚项，防止模型过拟合，改善模型的泛化能力。The beneficial effects of the invention: through the method, real-time monitoring of the state of the deep hole boring tool is realized, the dependence on the experience of the machine tool operator is reduced, the processing efficiency is improved, and the rejection rate is reduced. By filtering the original data to remove redundant data, speed up the training speed of the stacked autoencoder, and improve the accuracy of prediction; add a sparse penalty item to the loss function to prevent the model from overfitting and improve the generalization ability of the model.
图1为深孔镗刀状态监测流程图。Figure 1 is a flow chart of deep hole boring tool status monitoring.
图2为深孔镗床传感器布置示意图。Figure 2 is a schematic diagram of the sensor layout of a deep hole boring machine.
图3为第一层自编码器结构图。Figure 3 is a structural diagram of the first-layer self-encoder.
图4为堆叠自编码器结构图。Figure 4 is a structural diagram of a stacked self-encoder.
图5为断刀监测波形图。Figure 5 shows the broken tool monitoring waveform.
图6为磨钝监测波形图。Figure 6 is a blunt monitoring waveform diagram.
图中：1-工件；2-机床齿轮箱；3-传声器；4-床身；5-1#三向加速度传感器；6-刀杆；7-2#三向加速度传感器。In the picture: 1-workpiece; 2-machine tool gearbox; 3-microphone; 4-bed; 5-1# three-way acceleration sensor; 6-tool bar; 7-2# three-way acceleration sensor.
为了使本发明的技术方案和有益效果更加清晰明了，下面结合深孔镗刀状态监测的具体实施方式并参照附图，对本发明作详细说明。本实施例是以本发明的技术方案为前提进行的，给出了详细的实施方式和具体的操作过程，但本发明的保护范围不限于下述的实施例。In order to make the technical solutions and beneficial effects of the present invention clearer and clearer, the present invention will be described in detail below in conjunction with specific implementations of deep hole boring tool state monitoring and with reference to the accompanying drawings. This embodiment is performed on the premise of the technical solution of the present invention, and provides detailed implementation and specific operation procedures, but the protection scope of the present invention is not limited to the following embodiments.
以一台卧式深孔镗床加工深孔为例，详细说明本发明的实施方式。Taking a horizontal deep hole boring machine to process deep holes as an example, the implementation of the present invention will be described in detail.
第一步，深孔镗削过程中的振动和声音信息采集The first step is to collect vibration and sound information during the deep hole boring process
将传声器3固定在工件加工进口处，对准工件内孔。#1三向加速度传感器5和#2三向加速度传感器7用磁座吸附到刀杆保持架轴瓦侧面，安装方式如图2所示。设置采样频率1000Hz，采集加工过程中的振动和声信号。Fix the microphone 3 at the processing entrance of the workpiece and align it with the inner hole of the workpiece. The #1 three-way acceleration sensor 5 and #2 three-way acceleration sensor 7 are attached to the side of the tool holder bearing bush with a magnetic seat, and the installation method is shown in Figure 2. Set the sampling frequency to 1000 Hz to collect vibration and sound signals during processing.
第二步，振动和声音数据预处理The second step, vibration and sound data preprocessing
对采集到的数据做快速傅里叶变换，计算单侧幅值频谱。将频域数据，按照正常、断刀、磨钝三种状态进行分类，每个样本包含7000个数据。其中，正常状态下样本数量为20000，断刀和磨损状态下的样本数量分别为5000和3000。Perform fast Fourier transform on the collected data to calculate the one-sided amplitude spectrum. The frequency domain data is classified according to three states: normal, broken, and blunt. Each sample contains 7000 data. Among them, the number of samples in the normal state is 20000, and the number of samples in the broken and worn states are 5000 and 3000 respectively.
第三步，堆叠式自编码器网络的构建和训练The third step is the construction and training of the stacked autoencoder network
堆叠两自编码器和一个softmax分类器构建堆叠自编码器网络。第一个自编码器输入层和输出层神经元个数均为7000，隐层神经元个数为3000，如图3所示。第二个自编码器输入输出层神经元个数均为3000，隐层神经元个数为1000。第三个自编码器输入输出层神经元个数为1000，隐层神经元个数为500。随机选取三种状态下样本各2000个，对自编码器进行训练，采用贪婪训练法，对自编码器逐层进行训练。将两个自编码器和一个softmax分类器堆叠成堆叠自编码 器如图4所示，输出刀具的三种状态。再用剩余样本对网络进行测试，测试准确率达91.13％，模型可以用于刀具状态的监测。Stack two autoencoders and one softmax classifier to construct a stacked autoencoder network. The number of neurons in the input layer and output layer of the first autoencoder is 7000, and the number of neurons in the hidden layer is 3000, as shown in Figure 3. The number of neurons in the input and output layers of the second autoencoder is 3000, and the number of neurons in the hidden layer is 1000. The number of neurons in the input and output layer of the third autoencoder is 1000, and the number of neurons in the hidden layer is 500. Randomly select 2000 samples in each of the three states to train the autoencoder, and use the greedy training method to train the autoencoder layer by layer. Stack two autoencoders and a softmax classifier into a stacked autoencoder as shown in Figure 4, outputting three states of the tool. Then use the remaining samples to test the network, the test accuracy rate is 91.13%, and the model can be used to monitor the tool status.
第四步，深孔镗刀状态的实时监测The fourth step, real-time monitoring of deep hole boring tool status
将实时振动和声音数据经数据预处理后输入至堆叠自编码器模型中，对镗削加工过程中的刀具状态进行监测。如图5、6所示，模型能够准确地对刀具的实时状态做出判断。The real-time vibration and sound data are preprocessed and input into the stacked autoencoder model to monitor the tool status during the boring process. As shown in Figures 5 and 6, the model can accurately judge the real-time status of the tool.
Claims (1)
- 一种基于堆叠自编码器的深孔镗刀状态监测方法，其特征在于，步骤如下：A method for monitoring the state of a deep hole boring tool based on a stacked autoencoder, which is characterized in that the steps are as follows:第一步，深孔镗削过程中的振动和声音信息采集The first step is to collect vibration and sound information during the deep hole boring process将两个三向加速度传感器通过磁座吸附在深孔镗杆的两个保持架轴承上，将传声器放置于工件内孔的一端，对加工过程中的刀杆振动以及切削噪声进行采集；The two three-directional acceleration sensors are adsorbed on the two cage bearings of the deep hole boring bar through the magnetic base, and the microphone is placed at one end of the inner hole of the workpiece to collect the tool bar vibration and cutting noise during the machining process;第二步，振动和声音数据预处理The second step, vibration and sound data preprocessing对三向加速度传感器采集到的数据按照采样频率进行分段，设每段数据为x(n)，对每段数据做快速傅里叶变换：The data collected by the three-way acceleration sensor is segmented according to the sampling frequency, and each segment of data is set to x(n), and fast Fourier transform is performed on each segment of data:
其中，k＝0,...,N-1；Among them, k=0,...,N-1;计算其单侧幅频谱，将幅值低于0.2的数据全部滤掉；将滤波后的数据按照正常、磨钝和断刀三种状态进行分组；Calculate the one-sided amplitude spectrum, filter out all the data with amplitude less than 0.2; group the filtered data according to the three states of normal, blunt and broken;第三步，堆叠式自编码器网络的构建和训练The third step is the construction and training of the stacked autoencoder network取相同数量的三种状态下的样本数据做训练集，剩余样本做测试集，训练集和测试集要均包含三种状态下的样本数据；设输入数据为x＝[x 1,x 2,…,x i]，则自编码器的输入输出关系为： Take the same number of sample data in the three states as the training set, and the remaining samples as the test set. Both the training set and the test set must contain the sample data in the three states; set the input data as x=[x 1 ,x 2 , …,X i ], the input and output relationship of the autoencoder is:
式中，f(·)表示神经元的激活函数，表示输入层与隐含层的权重，
表示隐含层与输出层的权重，
示隐含层各神经元的偏置，
表示输出层各神经元的偏置； In the formula, f(·) represents the activation function of the neuron,
Represents the weight of the input layer and the hidden layer,Indicates the weight of the hidden layer and the output layer,Shows the bias of each neuron in the hidden layer,Represents the bias of each neuron in the output layer;构造损失函数：Construct the loss function:
式中，是自编码器的输出数据；α是稀疏惩罚项参数，作用是控制稀疏惩罚项在损失函数当中所占的比重；γ是稀疏性参数；
是隐含层第j个神经元的激活度； Where
Is the output data of the autoencoder; α is the sparse penalty term parameter, which is used to control the proportion of the sparse penalty term in the loss function; γ is the sparsity parameter;Is the activation degree of the jth neuron in the hidden layer;利用亚当优化器，最小化损失函数，优化权重w ij和偏置b： Using Adam optimizer, minimize the loss function, optimize the weight w ij and bias b:
式中，t为迭代次数；η为学习率；和
分别为一阶动量项和二阶动量项；λ 1和λ 2为动力值分别取0.9和0.999；
和
分别为一阶和二阶动量项的修正值；(w ij,b) t表示第t次迭代时的模型权重和偏置；g t＝ΔJ((w ij,b) t)表示t次迭代时代价函数关于(w ij,b) t的梯度；ε为常数，避免分母为0，取值很小； In the formula, t is the number of iterations; η is the learning rate;
withThey are the first-order momentum term and the second-order momentum term respectively; λ 1 and λ 2 are the dynamic values, taking 0.9 and 0.999 respectively;withRespectively are the correction values of the first-order and second-order momentum terms; (w ij ,b) t represents the model weight and bias at the tth iteration; g t =ΔJ((w ij ,b) t ) represents t iterations The gradient of the time cost function with respect to (w ij ,b) t ; ε is a constant, to avoid the denominator being 0, and the value is small;采用贪婪训练法对自编码器进行无监督训练，将训练集输入第一层自编码器，对其进行训练，最小化损失函数Loss，得到最优的权重和偏置；将第一层自编码器的隐层输出作为第二层自编码器的输入，对其进行训练，训练完成后用第二层自编码器的输出对softmax分类器，进行有监督训练，再将训练好的两层自编码器和softmax分类器堆叠起来，得到训练好的堆叠自编码器；训练完毕后，采用剩余测试集对堆叠自编码器进行测试；当测试准确率高于90％时，模型可用于刀具状态的监测；The greedy training method is used for unsupervised training of the autoencoder, the training set is input to the first layer of autoencoder, and it is trained to minimize the loss function Loss to obtain the optimal weight and bias; the first layer of autoencoder The hidden layer output of the second-layer autoencoder is used as the input of the second-layer autoencoder to train it. After the training is completed, the output of the second-layer autoencoder is used to perform supervised training on the softmax classifier, and then the trained two-layer autoencoder The encoder and the softmax classifier are stacked to obtain a trained stacked autoencoder; after training, the remaining test set is used to test the stacked autoencoder; when the test accuracy is higher than 90%, the model can be used for tool status monitor;第四步，深孔镗刀状态的实时监测The fourth step, real-time monitoring of deep hole boring tool status在实际加工过程中，将实时数据经数据预处理后输入至训练好的堆叠自编码器网络模型中，模型输出刀具的实时状态；当刀具状态正常时，模型输出为1；当刀具状态为断刀时，模型输出为2；当刀具状态为磨钝时，模型输出为3。In the actual machining process, the real-time data is input into the trained stacked autoencoder network model after data preprocessing, and the model outputs the real-time status of the tool; when the tool status is normal, the model output is 1; when the tool status is off When cutting, the model output is 2; when the tool is blunt, the model output is 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/CN2019/105283 WO2021046738A1 (en) | 2019-09-11 | 2019-09-11 | Method for monitoring state of deep hole boring cutter on basis of stacked autoencoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/CN2019/105283 WO2021046738A1 (en) | 2019-09-11 | 2019-09-11 | Method for monitoring state of deep hole boring cutter on basis of stacked autoencoder |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2021046738A1 true WO2021046738A1 (en) | 2021-03-18 |
Family
ID=74866875
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2019/105283 WO2021046738A1 (en) | 2019-09-11 | 2019-09-11 | Method for monitoring state of deep hole boring cutter on basis of stacked autoencoder |
Country Status (1)
| Country | Link |
|---|---|
| WO (1) | WO2021046738A1 (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016195358A1 (en) * | 2015-05-29 | 2016-12-08 | 한국생산기술연구원 | Method for processing carbon fiber reinforced plastic stack by using monitoring sensor |
| CN109015111A (en) * | 2018-07-06 | 2018-12-18 | 华中科技大学 | A kind of cutting tool state on-line monitoring method based on information fusion and support vector machines |
| CN109434563A (en) * | 2018-11-19 | 2019-03-08 | 深圳市蓝海永兴实业有限公司 | Numerical control cutting tool attrition real-time detecting system |
| CN109514349A (en) * | 2018-11-12 | 2019-03-26 | 西安交通大学 | Monitoring Tool Wear States in Turning based on vibration signal and Stacking integrated model |
| CN109562500A (en) * | 2016-08-10 | 2019-04-02 | 三菱重工工作机械株式会社 | The abnormal detector and method of the tool of lathe |
| JP2019098515A (en) * | 2017-11-28 | 2019-06-24 | 先馳精密儀器（東莞）有限公司Ｔｅｃｈｍａｒｋ Ｐｒｅｃｉｓｉｏｎ Ｉｎｓｔｒｕｍｅｎｔ Ｃｏ．Ｌｔｄ． | Blade tool state inspection system and method |
-
2019
- 2019-09-11 WO PCT/CN2019/105283 patent/WO2021046738A1/en active Application Filing
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016195358A1 (en) * | 2015-05-29 | 2016-12-08 | 한국생산기술연구원 | Method for processing carbon fiber reinforced plastic stack by using monitoring sensor |
| CN109562500A (en) * | 2016-08-10 | 2019-04-02 | 三菱重工工作机械株式会社 | The abnormal detector and method of the tool of lathe |
| JP2019098515A (en) * | 2017-11-28 | 2019-06-24 | 先馳精密儀器（東莞）有限公司Ｔｅｃｈｍａｒｋ Ｐｒｅｃｉｓｉｏｎ Ｉｎｓｔｒｕｍｅｎｔ Ｃｏ．Ｌｔｄ． | Blade tool state inspection system and method |
| CN109015111A (en) * | 2018-07-06 | 2018-12-18 | 华中科技大学 | A kind of cutting tool state on-line monitoring method based on information fusion and support vector machines |
| CN109514349A (en) * | 2018-11-12 | 2019-03-26 | 西安交通大学 | Monitoring Tool Wear States in Turning based on vibration signal and Stacking integrated model |
| CN109434563A (en) * | 2018-11-19 | 2019-03-08 | 深圳市蓝海永兴实业有限公司 | Numerical control cutting tool attrition real-time detecting system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110561192A (en) | Deep hole boring cutter state monitoring method based on stacking self-encoder | |
| CN111716150B (en) | Evolution learning method for intelligently monitoring cutter state | |
| CN108444708B (en) | Method for establishing rolling bearing intelligent diagnosis model based on convolutional neural network | |
| CN111136509B (en) | Cutter state monitoring method based on ELM-SDAE algorithm | |
| CN110355608A (en) | Based on the tool abrasion prediction technique from attention mechanism and deep learning | |
| CN105834834B (en) | Cutter wear state monitoring method based on drosophila optimized algorithm | |
| CN106112697A (en) | A kind of milling parameter automatic alarm threshold setting method based on 3 σ criterions | |
| Patange et al. | Milling cutter condition monitoring using machine learning approach | |
| CN109158954B (en) | Ultrasonic cutter wear state identification method and system based on acoustic signal detection | |
| CN110263474A (en) | A kind of cutter life real-time predicting method of numerically-controlled machine tool | |
| CN109262368A (en) | A kind of tool failure determination method | |
| CN109590805B (en) | A kind of determination method and system of turning cutting tool working condition | |
| CN114619292B (en) | Milling cutter wear monitoring method based on fusion of wavelet denoising and attention mechanism with GRU network | |
| Gong et al. | A fast anomaly diagnosis approach based on modified CNN and multisensor data fusion | |
| CN112247673A (en) | Woodworker cutter wear state diagnosis method based on genetic BP neural network | |
| WO2021046737A1 (en) | Method for monitoring tool wear in deep hole machining based on ssae-lstm model | |
| WO2021046738A1 (en) | Method for monitoring state of deep hole boring cutter on basis of stacked autoencoder | |
| CN114227382A (en) | Cutter damage monitoring system and method based on novel capsule network | |
| CN113043073A (en) | Cutter abrasion and service life prediction method and device | |
| CN108427375B (en) | Method for monitoring cutter state based on band-pass filtering processing multi-sensor | |
| CN115293211A (en) | Spindle state evaluation method and system based on multi-source information deep coupling network | |
| Li et al. | Deep learning based end-to-end rolling bearing fault diagnosis | |
| CN108956783A (en) | A kind of grinding sound wheel loading condition detection method based on HDP-HSMM | |
| Elbestawi et al. | Tool condition monitoring in machining | |
| Gui et al. | Unsupervised detection of tool breakage: A novel approach based on time and sensor domain data analysis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19945192 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 19945192 Country of ref document: EP Kind code of ref document: A1 |