JP2021074821A - Machine tool and display device - Google Patents

Abstract

To provide a machine tool that can support determination of a state during machining.SOLUTION: A machine tool 100 comprises: a detection part 101 that detects a detection value of at least one of energy supplied for machining, vibrations, sound, heat, light during machining and power generated during the machining; a feature quantity extraction part 102 that extracts a first feature quantity and a second feature quantity from the detection value detected by the detection part; and a display part 103 that displays, when a numerical value relating to the first feature quantity is set as a first axis and a numerical value relating to the second feature quantity is set as a second axis, a screen including at least two border lines arranged in a contour line shape in order to represent a possibility that an abnormality occurs during the machining and a point having the detection value plotted.SELECTED DRAWING: Figure 1

Description

The present invention relates to machine tools and display devices.

In the above technical field, in Patent Document 1, a learning model is constructed by performing machine learning using dynamic information such as the scattering direction, scattering range, and scattering speed of chips near the tool generated by machining, and the learning model is used. A technique for estimating tool wear or loss is disclosed.

Japanese Unexamined Patent Publication No. 2018-0138327

However, the technique described in the above document has not been able to support the determination of the state at the time of processing.

An object of the present invention is to provide a technique for solving the above-mentioned problems.

In order to achieve the above object, the machine tool according to the present invention
A detector that detects at least one of the energy, vibration, sound, heat, light, and power generated during processing during processing.
A feature amount extraction unit that extracts a first feature amount and a second feature amount from the detected value detected by the detection unit, and a feature amount extraction unit.
The point where the detected value is plotted on a plane with the numerical value related to the first feature amount as the first axis and the numerical value related to the second feature amount as the second axis, and the possibility that an abnormality may occur during the processing. It is provided with at least two boundary lines arranged in a contour line to represent, and a display unit for displaying.

In order to achieve the above object, the display device according to the present invention is
Detected from a machine tool having a detection unit that detects a detection value of at least one of energy, vibration, sound, heat, light, and power generated during processing during processing. A display device that extracts the first feature amount and the second feature amount from the detected values and displays the machining status of the machine tool.
The point where the detected value is plotted on a plane with the numerical value related to the first feature amount as the first axis and the numerical value related to the second feature amount as the second axis, and the possibility that an abnormality may occur during the processing. At least two borders arranged in contour lines to represent.

According to the present invention, it is possible to support the determination of the state at the time of processing.

It is a figure which shows the structure of the machine tool which concerns on 1st Embodiment. It is a figure which shows the appearance of the machine tool which concerns on 2nd Embodiment, and the schematic structure of the tool holding part. It is a figure which shows an example of the one-dimensional graph by the premise technique of the machine tool which concerns on 2nd Embodiment. It is a figure which shows an example of the one-dimensional graph by the premise technique of the machine tool which concerns on 2nd Embodiment. It is a figure which shows the internal structure of the machine tool which concerns on 2nd Embodiment. It is a figure explaining the state of change of the 2D map before and after the correction by the correction part of the machine tool which concerns on 2nd Embodiment. It is a figure for demonstrating the extraction of the feature amount by the feature amount extraction part of the machine tool which concerns on 2nd Embodiment. It is a figure for demonstrating the locus of a two-dimensional map. It is a figure which shows an example of the mapping target table which the machine tool which concerns on 2nd Embodiment has. It is a flowchart explaining the processing procedure of the machine tool which concerns on 2nd Embodiment. It is a figure explaining the structure of the machine tool which concerns on 3rd Embodiment. It is a figure explaining the biting data deletion by the machine tool which concerns on 3rd Embodiment.

Hereinafter, embodiments for carrying out the present invention will be described in detail, exemplarily, with reference to the drawings. However, the configuration, numerical values, processing flow, functional elements, etc. described in the following embodiments are merely examples, and modifications and changes thereof are free, and the technical scope of the present invention is described below. It is not intended to be limited.

[First Embodiment]
The machine tool 100 according to the first embodiment will be described with reference to FIG. FIG. 1 is a diagram for explaining the configuration of the machine tool 100 according to the present embodiment.

The machine tool 100 includes a detection unit 101, a feature amount extraction unit 102, and a display unit 103. The detection unit 101 detects the detection value of at least one of energy, vibration, sound, heat, light, and power generated during processing during processing. The feature amount extraction unit 102 extracts the first feature amount and the second feature amount from the detection value detected by the detection unit 101. The display unit 103 plots the detected values on a plane having the numerical value related to the first feature amount as the first axis 134 and the numerical value related to the second feature amount as the second axis 135 (T1 to T16 in the figure). And at least two boundary lines 132 and 133 arranged in a contour line to indicate the possibility of an abnormality occurring during processing.

The points (T1 to T10 in the figure) displayed inside the boundary line 132 displayed on the display unit 103 indicate that the machine tool 100 is operating normally. In addition, the points (T11 to T14) displayed between the boundary line 132 and the boundary line 133 are operating normally, but a minor abnormality (abnormality considered to be a sign of breakage) that does not affect the processing occurs. Indicates that there is a possibility of doing so. Further, the points (T15 to T16) displayed on the outside of the boundary line 133 indicate that abnormalities affecting the machining (tool breakage / chipping, damage to the machining work) may occur.

The machine tool 100 in the present invention may be an additive manufacturing machine that processes by adhering a material, a subtractive manufacturing machine that removes the material, or a machine that irradiates light such as a laser. Good. Specifically, lathes, drilling machines, boring machines, milling machines, gear cutting machines, grinding machines, multi-axis machines, laser machines, laminating machines and the like are included. These perform various processes such as turning, cutting, drilling, grinding, polishing, rolling, forging, bending, molding, microfabrication, and lathe processing on workpieces such as metal, wood, stone, and resin. .. A multifunction device that combines these processes is also included.

According to this embodiment, since the machining state is displayed as a two-dimensional map, it is possible to support the state determination at the time of machining. In addition, since the state during machining can be accurately determined, it is possible to prevent breakage of the tool and the like, and for example, it is possible to prevent a decrease in productivity such as a long machining time due to the breakage of the tool.

[Second Embodiment]
Next, the machine tool 200 according to the second embodiment will be described with reference to FIGS. 2 to 7. FIG. 2 is a diagram showing the appearance of the machine tool and the schematic configuration of the tool holding portion according to the present embodiment. The machine tool 200 of the second embodiment will be described using a multifunction device. The machine tool 200 has a tool holding portion 201, and the tool 202 is attached to the tool holding portion 201. The tool 202 processes the work 203 placed on the table 204. An object of the present embodiment is to prevent the tool 202 from being broken in advance.

FIG. 3A is a diagram showing an example of a one-dimensional graph as a prerequisite technique of the present embodiment. In FIG. 3A, a graph 351 in which the current value applied to the tool spindle is set on the vertical axis 356 and the times T1 to T16 are set on the horizontal axis 357 is displayed. When using such a graph, assuming that the breakage timing is T16, the timing of T15 is actually the appropriate alert timing. However, when trying to detect the timing of T15, it is necessary to set a threshold value 358 in advance. Then, the replacement of the tool 202 is urged at the timing of T9 when the threshold value 358 is exceeded. Therefore, although the tool can still be used, the tool 202 has to be replaced, and waste is generated for the time of T9 to T15. That is, since the frequency of tool replacement increases, the overall machining time becomes long, and the tool cost per machining time increases. Further, as shown in FIG. 3B, there is a problem that it is difficult to discriminate by the threshold method with a small-diameter tool, and it is easy to make a mistake in the timing of tool replacement. For example, the current value became about 20A just by rotating the spindle, and the variation was large. Even if the current value was measured by rotating a tool of φ3.3, it was not possible to distinguish whether it was cutting or not. That is, the simple threshold method could not determine the sign of breakage of a small-diameter tool. The lower one-dimensional graph in FIG. 3B is a partially enlarged view of the upper one-dimensional graph.

FIG. 3C is a diagram showing an internal configuration of the machine tool 200. The machine tool 200 includes a detection unit 301, a feature amount extraction unit 302, a display unit 303, a correction unit 304, an operation unit 305, a boundary data holding unit 306, an abnormality determination unit 308, and a boundary data generation unit 309. The machine tool 200 displays a two-dimensional map 331 for determining the possibility of tool breakage on the display unit 303 based on the detection value detected by the detection unit 301.

The detection unit 301 detects the current value applied to the tool spindle to rotate the tool 202 as an example of the energy applied to the machine tool 200 for machining, and outputs it as a detected value (sensed value). The detection unit 301 includes a current sensor provided in the three-phase alternating current UVW phase and an AD converter that converts the current value detected by the current sensor into digital data. For example, with an AD converter, the sampling frequency is set to 2 kHz, 16 bits, that is, 256 points of time series data can be acquired and used as input data every 128 msec.

上の式において、Ｑ軸電流ｉ_qは実効電流、Ｄ軸電流ｉ_dは無効電流である。検知部３０１は、Ｑ軸電流ｉ_qを検知値として特徴量抽出部３０２に送る。 The detection unit 301 calculates the _{Q-axis current i q} and the D-axis current i _d with respect to the digital current value output from the AD converter using the following conversion formula (1).

In the above equation, the Q-axis current i _q is the effective current and the D-axis current i _d is the reactive current. The detection unit 301 sends the Q-axis current i _q as a detection value to the feature amount extraction unit 302.

The feature amount extraction unit 302 has a frequency decomposition unit 321 and a normalization unit 322 and a dimension compression unit 323. The frequency decomposition unit 321 extracts a frequency component from the detected value received from the detection unit 301 by using, for example, a Fourier transform. The normalization unit 322 normalizes the frequency component data after the frequency decomposition. The dimensional compression unit 323 compresses the dimension of the normalized frequency component data to generate a two-dimensional feature amount (data having a first feature amount component and a second feature amount component). The feature amount extraction unit 302 is a processor for executing a predetermined program.

The display unit 303 displays a two-dimensional map 331 showing the possibility of tool breakage based on the two-dimensional feature amount data extracted by the dimensional compression unit 323. The two-dimensional map 331 includes a plane in which the first feature amount generated by the dimensional compression unit 323 is the first axis 332 and the second feature amount is the second axis 333. The features of the detected values are plotted on the plane (T1 to T16). Further, the display unit 303 displays on the screen the boundary lines (three boundary lines 334 to 336 in this example) arranged in contour lines in order to show the difference in the possibility that an abnormality occurs during processing. In the present embodiment, "abnormality" means breakage of the tool.

The two-dimensional map 331 shows that the farther the feature amount of the plotted detection values is from the center of the innermost boundary line 334 to the outside, the higher the possibility that an abnormality will occur.

For example, the feature quantities (T1, T8, T9 ...) Of the detected values plotted inside the boundary line 334 can be judged to be normal because the possibility of occurrence of an abnormality is extremely low. That is, the user who sees the two-dimensional map 331 can continue the processing with peace of mind as long as the point representing the feature amount of the detected value is displayed only inside the boundary line 334.

If the point representing the feature amount of the detected value is between the boundary line 334 and the boundary line 335, the user may judge that the abnormality is unlikely to occur but is equal to or more than the predetermined value, and carefully process the value. For example, the rotation speed of the tool 202 can be set to a rotation speed at which breakage is unlikely to occur, and machining can be continued. Further, if there is a feature amount of the detected value between the boundary line 334 and the boundary line 335, the user should start considering replacement of the tool 202 and the like.

For example, referring to the two-dimensional map 331, since T11 to T14 are points outside the boundary line 334, there is a high possibility that an abnormality will occur in processing. Since the point T15 is outside the boundary line 335, it can be seen that the tool 202 needs to be replaced immediately. By displaying the two-dimensional map 331 in this way, the timing of replacement of the tool 202 can be determined more accurately than the one-dimensional graph 351.

In addition, by looking at the locus of points, it is possible to easily grasp signs of abnormalities that occur in the machine tool 200 over the long term, so that the user can easily grasp the medium- to long-term maintenance plan of the machine tool 200 and the procurement of consumables. You can make a plan.

Further, for example, if the feature amount of the detected value is plotted outside the boundary line 336, it is considered that the tool 202 is likely to be broken immediately, so that the tool 202 should be replaced promptly. Is.

Each time the detection unit 301 detects a detected value, the display unit 303 additionally displays a point on which the characteristics of the detected value are plotted, and at the same time displays a boundary line that serves as a criterion for determining normality or abnormality.

The display unit 303 may be, for example, a display provided as a part of the machine tool 200, or a display outside the machine tool 200. Alternatively, a projector may be used to project the two-dimensional map 331 onto the screen. In this case, the display device is a machine based on a feature quantity extracted from at least one of vibration, sound, light, current loaded on the tool 202, heat, and power value detected as detected values during machining. The processing state is displayed as a two-dimensional map 331. Information about the two-dimensional map 331 may be held by the display device side.

The correction unit 304 corrects the two-dimensional map displayed by the display unit 303 according to an instruction from the user. Specifically, as shown in FIG. 3D, the correction unit 304 corrects the shape of the boundary line 336 of the two-dimensional map 331 in the direction of expanding to the lower left side. The user can widen the range of normal detection values by widening the boundary line 334. By correcting the shape of the boundary line 334 by a veteran user who can determine normality or abnormality, other users who use the machine can follow the judgment of the veteran user. FIG. 3D shows an example of correction so as to widen the boundary line 334 so that the points T12 to T14 fit inside the boundary line 334.

There is a difference in the boundary lines 334 to 336 of the two-dimensional map 331 drawn between the detection value acquired at the time of shipment of the machine tool 200 from the factory and the detection value acquired when the machine tool 200 is placed in the user's factory or the like. In some cases. That is, the machining conditions by the machine tool 200 are not always the same.

Since the hardness of the work material is different for the boundary lines 334 to 336, for example, the shape of the boundary lines 334 to 336 of the two-dimensional map 331 showing the possibility of breaking of the tool differs depending on which work material is used. .. The tool also behaves differently for each work. Therefore, the two-dimensional map can be displayed according to the user's usage environment by configuring the shape of the boundary lines 334 to 336 of the drawn two-dimensional map 331 to be corrected.

As a correction method, for example, the user may change the shape of the boundary lines 334 to 336 by dragging a part of the boundary lines 334 to 336 using an operation unit 305 such as a mouse. Further, the shape of the boundary lines 334 to 336 may be changed by inputting a numerical value using an operation unit 305 such as a keyboard. In addition, data on the corrected boundary line may be saved in the cloud and shared with other machine tools.

The abnormality determination unit 308 determines whether or not there is an abnormality based on the characteristics of the detection value extracted by the feature amount extraction unit 302. The abnormality determination unit 308 passes the determination result to the boundary data generation unit 309.

The boundary data generation unit 309 generates a boundary line between the point determined to be normal and the point determined to be abnormal in the abnormality determination unit 308.

The boundary data holding unit 306 holds the data of the boundary lines 334 to 336 of the two-dimensional map 331 displayed on the display unit 303. The correction unit 304 corrects the shape of the boundary lines 334 to 336 by changing the data of the boundary lines 334 to 336 held by the boundary data holding unit 306.

(Feature amount extraction process)
Here, with reference to FIG. 3E, a method of extracting the feature amount by the feature amount extraction unit 302 will be described in detail. The frequency decomposition unit (FFT in the figure) 321 extracts the frequency component of the current value detected by the detection unit 301 during machining by the machine tool 200 to generate a frequency spectrum. The frequency decomposition is performed by, for example, FFT (Fast Fourier Transform), but is not limited thereto.

In general, any periodic time series data y _t can be thought of as the sum of trigonometric functions of various periods. This is called a Fourier series expansion, and the Fourier series of the observed value y _{t whose basic period is T 0} [s] can be expressed by Eq. (2) using a complex number.

Here, ω ₀ is a fundamental angular frequency and is _{represented by ω 0} = 2πf ₀ [rad / s]. f ₀ is the fundamental frequency and is _{represented by f 0} = 1 / T ₀ [Hz].

On the other hand, the complex Fourier coefficient c _n can be obtained by the following equation (3).

When the sampling time of the current data is d and the window length is n, the maximum frequency f _max and the frequency resolution Δf are expressed by the following equations (4) and (5).

For example, when the sampling frequency is 2 kHz and the window length is 256 points, the maximum frequency f _max after FFT is 1 kHz, and the frequency resolution Δf is 7.8125 Hz. That is, the 256-point time-series data can be represented by a vector of 128 points when FFT is performed, and this vector becomes an input to the next autoencoder (AE in the figure). The 128-point vector is normalized by the normalization unit 322 and sent to the dimension compression unit 323.

The dimensional compression unit 323 performs dimensional compression using an autoencoder 361 and a PCA (Principal Component Analysis) 362. The autoencoder 361 is an algorithm for dimension compression using a neural network in machine learning, and is an algorithm capable of extracting features having an overwhelmingly smaller number of dimensions than the number of dimensions of an input sample.

In the present embodiment, the spindle current value is frequency-decomposed by the frequency decomposition unit 321 into data represented by a multidimensional vector (128 dimensions in this case). Then, the frequency-decomposed multidimensional vector (128 dimensions) is used as the input of the dimensional compression unit 323. By setting the intermediate layer of the autoencoder 361 of the dimensional compression unit to a low dimension, the vector input of a plurality of dimensions is dimensionally compressed to a low dimension. In general, an autoencoder 361 uses the same data for an input layer and an output layer in a three-layer neural network, compresses the input layer to an intermediate layer, and then repeatedly restores the output layer, and has a high recall rate. It derives the middle layer. Here, the intermediate layer is set to 64 dimensions. That is, the 128-dimensional vector input to the dimensional compression unit 323 is compressed to 64 dimensions while maintaining the features as much as possible. For example, when the vector after FFT is 128 dimensions, it can be made 64 dimensions or 10 dimensions by the autoencoder 361. The learning of the autoencoder 361 and the processing using the trained model will be described below.

ここで、

は、実験番号βのｒ番目の区間のベクトルを表わしており、Ｒは区間総数である。 (I) FFT data x ^beta of the spindle current waveform learning experiment numbers beta of Autoencoder is represented by a set of R-dimensional vector by the following equation (6).

here,

Represents the vector of the r-th interval of the experiment number β, and R is the total number of intervals.

また、デコーダ部は、以下の式（８）で表わされる。

ここで、Ｗ’は、Ｗの転置行列であるため、求めるパラメータは、Ｗ、ｂ、ｂ’の３つとなる。 Then, the encoder unit is represented by the following equation (7).

The decoder unit is represented by the following equation (8).

Here, since W'is a transposed matrix of W, the parameters to be obtained are W, b, and b'.

は、元信号

に類似したものとなる。 in this way,

Is the original signal

Will be similar to.

ここで、β₁、β₂∈［０，１）はハイパーパラメータであり、例えば、Ａｄａｍの推奨値である下記の値βを用いてもよいが、推奨値を基準に調整してもよい。 For example, the use of Adam (Adaptive moment estimation) the optimization technique, the square of the moving average of the slope up to the step t-1 of the immediately preceding ^{_{ν t = E [g 2]}} t and slope moving average m _t = E of [ g] _t can be expressed as the following equation (9).

Here, β ₁ , β ₂ ∈ [0, 1) is a hyperparameter, and for example, the following value β, which is a recommended value of Adam, may be used, but may be adjusted based on the recommended value.

β ₁ = 0.9
β ₂ = 0.999

Here, assuming _{that initialization is performed with ν 0} = 0, the following equation (10) is obtained.

ここで、ζ＝０と近似できるようにハイパーパラメータの値を設定すれば、

が求められる。以上より、

がパラメータ更新式となる。バイアスのｂ、ｂ’においても同様の手順で導出することが可能である。 In other words, the relationship of the moving average E of the second moment ν _t [ν _t] the true second moment E [g ² _t] is expressed by the following equation (11).

Here, if the value of the hyperparameter is set so that it can be approximated as ζ = 0,

Is required. From the above,

Is the parameter update formula. Bias b and b'can be derived by the same procedure.

例えば、ＦＦＴ後のベクトルが１２８次元の場合で、中間層を６４次元に設定してオートエンコーダを学習させた場合、１２８次元の入力

から次元圧縮された６４次元の

が出力される。 (Ii) Autoencoder processing The learned autoencoder encoder is used for dimensional compression.

For example, when the vector after FFT is 128 dimensions and the intermediate layer is set to 64 dimensions and the autoencoder is trained, the input is 128 dimensions.

64 dimensional compressed from

Is output.

PCA (Principal Component Analysis) is a method of expressing data in a lower dimension using principal components. It is possible to compress the data of any number of input dimensions and output it.

と表される。この式（１５）を行列とベクトルとの形式で書き直すと、

となる。つまり、ｗのノルムが１となる拘束条件を満たした上で、出力Ｘｗが最大になるｗを求める。その答えはＸの特異値分解にある。ｋ次元の主成分をＶ_k、ｗ、データ行列をＷとすると、
Ｗ＝Ｖ_ｋ
である。ここで、ｋは、主成分Ｖに含まれる相関のない変数群の個数である。よって、次元圧縮後のベクトルをＺとすると、射影座標のベクトルは、
Ｚ＝ＸＶ_k
と表される。 (I) Let X be the data matrix of the learning input, and x be the column vector containing a single data point. When we formulate the data distribution maximization problem,

It is expressed as. Rewriting this equation (15) in the form of a matrix and a vector,

Will be. That is, after satisfying the constraint condition that the norm of w is 1, the w that maximizes the output Xw is obtained. The answer lies in the singular value decomposition of X. _{Let V k} and w be the principal components of the k dimension and W be the data matrix.
W = V _k
Is. Here, k is the number of uncorrelated variable groups included in the principal component V. Therefore, assuming that the vector after dimensional compression is Z, the vector of the projected coordinates is
Z = XV _k
It is expressed as.

(Ii) Processing Dimension compression is performed using the principal component V obtained by learning. For example, when the vector dimension after the autoencoder processing is 64 dimensions, the 64-dimensional input X ₆₄ can be set to be dimensionally compressed into the two-dimensional vector Z _2.

Z ₂ = X ₆₄ V ₂
Therefore, when a 64-dimensional X ₆₄ is input, the two-dimensional vector Z ₂ is output by the _{principal component V 2.} The two vector components of the vector Z ₂ correspond to the first feature quantity and the second feature quantity. The two components of the vector Z ₂ are plotted on a two-dimensional map as the first and second axes.

The SVM (Support Vector Machine) 363 is a pattern recognition method originally intended for two-class classification, and the SVM 363 seeks an optimum separation hyperplane that maximizes the margin. Here, the margin is the distance between the sample closest to the separation hyperplane and the separation hyperplane. The maximized margin (distance) is represented by f (x).

The SVM363 is a One Class SVM (One Class SVM), which is an extension of a normal SVM, and constructs a model that maps normal data to a non-negative value and abnormal data to a negative value. That is, One Class SVM is a method of performing learning based on a set of data in which the majority is normal, and determining whether the unknown data is normal or abnormal. In general, although a lot of data on normally made products and normal conditions can be obtained, data on abnormal products and abnormal conditions cannot be obtained very much. It is possible to apply One Class SVM to such cases.

ここで、ｘ_iが第１特徴量であり、ｘ_jが第２特徴量である。また、φは非線形関数である。ＳＶＭモデルによって決まる距離ｆ（ｘ）は以下の式（１８）で表される。

(I) Learning When the model construction data cannot be linearly separated, SVM uses a nonlinear function to map the model construction data to a high-dimensional space and obtain a separation hyperplane in the high-dimensional space. This is equivalent to finding a non-linear separation boundary in the original low-dimensional space. It is mapped to a higher dimensional space using the kernel function K.

Here, x _i is the first feature quantity and x _j is the second feature quantity. Also, φ is a non-linear function. The distance f (x) determined by the SVM model is expressed by the following equation (18).

Here, w is a weight vector and b is a bias. Here, the new sample x can be identified by the sign of f (x). Model construction in One Class SVM is formulated as SVM when normal data for model construction is regarded as the same class and the origin is regarded as the other class. That is, the margin in One Class SVM is defined as the distance between the origin and the sample closest to the origin, and the margin maximization problem is formulated as in the following equation (19).

Ｌａｇｒａｎｇｅの未定乗数法により、以下の式（２１）を導くことができる。

整理すると、この問題は以下のような双対形式の式（２２）で表現できる。

Here, ξ _i (i = 1, ..., N) is a slack variable, and ν ∈ (0, 1) is the error rate when the model construction data is identified by the constructed One Class SVM model. .. Here, when the Lagrange multipliers α _i ≧ 0 and η _i ≧ 0 are introduced, this optimization problem is rewritten as the following equation (20).

The following equation (21) can be derived by Lagrange's undetermined multiplier method.

To summarize, this problem can be expressed by the following dual-form equation (22).

通常、異常検知モデルのような識別モデルを構築するには、正常データと異常データとの両方が必要である。しかし、Ｏｎｅ Ｃｌａｓｓ ＳＶＭでは、上記式（２３）で分かるとおり、正常データのみから識別モデルを構築できる。 This optimization problem can be solved as a standard quadratic programming problem. The distance f (x) finally determined by the One Class SVM model is expressed by the following equation (23).

Usually, both normal data and anomaly data are required to build a discriminative model such as an anomaly detection model. However, in One Class SVM, as can be seen from the above equation (23), the discriminative model can be constructed only from normal data.

As described above, the One Class SVM outputs the distance f (x) from the hyperplane. Normal or abnormal can be determined based on the distance f (x). The distance f (x) has a negatively large value for abnormal data and a positively large value for normal data.

Here, the boundary data generation unit 309 uses data having a small distance f (x), that is, abnormal data, as outliers to generate a boundary line.

For example, assuming that the training data contains an abnormal value of 0.2%, the abnormality determination unit 308 creates a normal range by setting the upper 0.2% as outliers in ascending order of the distance f (x). can do.

(Ii) Processing By inputting the dimensionally compressed data by PCA into the trained One Class SVM, the output f (x) can be obtained from the above equation (23).

The abnormality determination unit 308 uses the output f (x) whose positive and negative are reversed as the abnormality score g (x).
g (x) = −f (x)
It is expressed as. That is, the abnormality determination unit 308 determines that the abnormality score g (x) is normal when it is a negative value, and that it is abnormal when it is 0 or more.

The boundary data generation unit 309 generates a boundary line between the points determined to be normal by the abnormality score g (x) and the points determined to be abnormal, and stores the boundary data as boundary data in the boundary data holding unit 306. To do.

In this way, in the present embodiment, the dimension of the detected detection value is once compressed (128 dimension → 64 dimension → 2 dimension) to make it easier to judge whether it is normal or abnormal, and a line between normal and abnormal is drawn. ing. Of course, it is not limited to this, and One Class SVM is used to judge whether it is normal or abnormal based on 128-dimensional data, and separately, dimension compression (128-dimensional → 64-dimensional → 2-dimensional). And plot on a two-dimensional plane. In this case, although it takes a long time to process, the determination of whether it is normal or abnormal is more accurate than using two-dimensional data.

Further, although PCA362 is used as the dimensional compression method, for example, VAE (Variational AutoEncoder) may be used instead of PCA362. Further, the compression may be performed up to two dimensions using only the autoencoder 361 without using the PCA 362.

The dimensional compression method is not limited to the method shown here, and various methods may be used in combination. Further, the example shown here is an example of creating a two-dimensional map for the breakage of the tool 202, but for the chattering of the tool spindle, only the first detected value is different, and the subsequent processing is performed. It can be done in the same way, and a similar two-dimensional map can be created. For example, when it is desired to detect chatter, data of sound or vibration generated during machining may be acquired as a detection value.

Here, chatter is vibration that is continuously generated between the tool and the work piece. When chatter occurs, traces (chatter marks) corresponding to chatter remain on the work piece, and the finished surface deteriorates. Further, at this time, an abnormal vibration noise may be generated, which may damage the tool or the holder holding the tool, or damage a part of the machine tool. Chatter is roughly divided into forced chatter and self-encouragement chatter according to the cause of occurrence. Forced chatter occurs when there is a vibration source somewhere in the cutting system and it is affected by it. Vibration sources include vibration of the drive device inside the machine tool, intermittent cutting force in milling, and vibration transmitted from the outside of the machine tool. On the other hand, self-excited chatter occurs when a certain condition is satisfied in which the dynamic characteristics of the machine tool and the cutting process overlap even if there is no specific vibration source. For example, it tends to occur during heavy cutting and high-speed cutting.

FIG. 4 is a diagram for explaining the locus of points in the two-dimensional map 331. Although points exist within the normal range in any of the upper, middle, and lower two-dimensional maps 331, the locus 431 gradually moves away from the center of the boundary line as compared with the loci 411 and 421. Therefore, caution is required.

That is, according to the two-dimensional map 331, it is possible to more accurately grasp the change in the processing state as compared with the conventional one-dimensional display.

FIG. 5 is a diagram showing an example of a table 501 included in the machine tool 200 according to the present embodiment. Table 501 stores the detection value 512 in association with the mapping target 511 to be mapped as a two-dimensional map. The current, sound, vibration, torque, etc. applied to the spindle are stored as the detection values 512 to be acquired and analyzed for each of the mapping target 511, the tool breakage, the spindle chatter, and the cutting state. The machine tool 200 refers to the table 501 to determine the data to be acquired in order to display the two-dimensional map 331.

The machine tool 200 as described above has a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), and a storage as hardware. The machine tool 200 reads the data necessary for realizing the present embodiment into the RAM and executes it by the CPU. The storage stores databases, various parameters, data, programs, and modules.

FIG. 6 is a flowchart illustrating a processing procedure of the machine tool 200 according to the present embodiment. When the CPU executes the program according to this flowchart, each functional configuration shown in FIG. 3C is realized.

In step S601, the detection unit 301 detects at least one of the applied current value, vibration, sound, heat, light, and power generated during processing as detection values from various sensors. Next, in step S603, the feature amount extraction unit 302 reduces the dimension of the feature amount of the detected detected value, and extracts the first feature amount and the second feature amount. In step S605, the display unit 303 generates a screen of the two-dimensional map 311 in which the detected values are plotted on a plane having the first feature amount as the first axis 332 and the second feature amount as the second axis 333. Further, the display unit 303 generates three boundary lines 334 to 336 arranged in contour lines to indicate the possibility that an abnormality may occur in processing.

In step S607, the display unit 303 displays the created two-dimensional map 331 on the display. In step S609, the correction unit 304 determines whether or not to correct the boundary line shape. When the boundary line shape is not corrected (NO in step S609), the machine tool 200 ends the process. When correcting the boundary line shape (YES in step S609), the process proceeds to step S611. In step S611, the display unit 303 displays the two-dimensional map 331 corrected for the boundary line shape.

In the description of this embodiment, the detection unit 301 has been described as an example of detecting the current value, but the detection unit 301 is not limited to the current value, and for example, vibration, sound, heat, and the like. It may be the respective values of light and power generated during processing.

According to the present embodiment, the user can intuitively grasp a sign of an abnormality during machining such as breakage of a tool or occurrence of chatter. As a result, the machine tool can be stopped and the tool 202 can be replaced before the tool 202 is broken, and the tool can be prevented from being broken. For example, it is possible to prevent a decrease in productivity such as a long machining time due to a broken tool.

Further, the user of the machine tool can adjust the rotation speed of the tool spindle while looking at the two-dimensional map. Furthermore, the rotation speed of the tool spindle that does not cause chatter can be adjusted. In addition, by accumulating the detected values, it is possible to predict future breakage of the spindle. In addition, the user of the machine tool can grasp the state of machining in real time. Even if there is no failure (unnormal) data that is usually difficult to obtain, if there is normal (success) data, a highly accurate prediction model can be generated, and highly accurate prediction becomes possible. Further, even if there is no normal data (teacher data), if there is normal data in the vicinity, a highly accurate prediction model can be generated using the normal data in the vicinity, and highly accurate prediction becomes possible. Furthermore, instead of unilaterally notifying the user of the sign or occurrence of an abnormality from the machine tool, there is room for the user to intuitively judge or think by looking at the two-dimensional map, and what causes the abnormality. The user can determine the reason for the notification.

Further, for example, consider that a user tests which combination of tools, tool spindles, workpieces, etc. is efficient for mass production of a certain processed product. By displaying a two-dimensional map with multiple boundaries for the test results, the user can load the combination of tools, tool spindles, workpieces, etc. from the positional relationship between the boundaries and points on the map. You can judge whether you can call. As a result, the user can mass-produce processed products efficiently and at low cost.

The display unit 303 may display a plurality of two-dimensional maps in parallel on one screen. For example, the display unit 303 may simultaneously display a two-dimensional map for tool breakage and a two-dimensional map for spindle chatter on one screen. By displaying a plurality of two-dimensional maps at the same time in this way, the possibility of occurrence of various abnormalities can be confirmed on one screen.

[Third Embodiment]
Next, the machine tool 700 according to the third embodiment of the present invention will be described with reference to FIGS. 7 and 8. The machine tool according to the present embodiment is different from the second embodiment in that it has a bite data deletion unit 701.

The bite data deletion unit 701 uses a Kalman filter to determine whether or not there is bite data in the detected value such as the detected current value. Then, when it is determined that there is bite data, the bite data deletion unit 701 deletes the data for several seconds from the processing start point of the detected detection value, and sends the detection value from which the bite data is deleted to the feature amount extraction unit 302. ..

The display unit 703 in the present embodiment is provided with a touch screen, and when the point "T10" of the displayed two-dimensional map 331 is touched (or clicked with a mouse or the like), information (date and time, detected value, tool type) related to the point T10 is obtained. , Usable time, processing conditions, etc.) are displayed. By selecting the points on which the detected values are plotted in this way, information related to the detected values can be displayed on the screen, and a more detailed processing state can be known. As a result, it is possible to grasp the processing state several seconds ago or several minutes ago. In addition, the machining conditions for the next machining can be selected with reference to the positions of points indicating the past machining states on the two-dimensional map, the detection values of the points, and the machining conditions.

FIG. 8 is a diagram illustrating the deletion of biting data by the machine tool according to the present embodiment. As shown in Graph 801 when a two-dimensional map is generated in a state where the data at the time of biting, which is the machining start part by the machine tool, is included, the detection value that is the source of the two-dimensional map generation is due to the bite. Noise is included. Therefore, it is not possible to generate a highly accurate two-dimensional map. Therefore, as shown in Graph 802, if the data for the first few seconds of the detected detected value is deleted and a two-dimensional map is generated, the two-dimensional map can be generated with the detected value that does not include noise. Can generate a high 2D map.

Here, at the beginning of the graph 801 there is a bite to the detected current value, and it can be seen that a large low frequency is generated. Therefore, the presence of this biting portion becomes a factor that hinders the detection of the presence or absence of an abnormality. Therefore, as shown in Graph 802, for example, data for 600 points (1/2000 × 600 = 0.3 [s]) is deleted from the machining start point by the machine tool. It should be noted that the determination as to whether or not the bite data is included in the graph 801 is performed using a Kalman filter.

According to the present embodiment, since the bite data is deleted when there is a bite portion, it is possible to generate a highly accurate two-dimensional map by using the detection value with less extra noise.

[Other Embodiments]
Although the invention of the present application has been described above with reference to the embodiment, the invention of the present application is not limited to the above embodiment. Various changes that can be understood by those skilled in the art can be made to the structure and details of the present invention within the technical scope of the present invention. Also included in the technical scope of the present invention are systems or devices in any combination of the different features contained in each embodiment.

Further, the present invention may be applied to a system composed of a plurality of devices, or may be applied to a single device. Further, the present invention is also applicable when an information processing program that realizes the functions of the embodiment is supplied to a system or an apparatus and executed by a built-in processor. Therefore, in order to realize the functions of the present invention on a computer, the program installed on the computer, the medium containing the program, the WWW (World Wide Web) server for downloading the program, and the processor for executing the program are also included in the present invention. Included in the technical scope of the invention. In particular, at least a non-transitory computer readable medium containing a program that causes a computer to execute the processing steps included in the above-described embodiment is included in the technical scope of the present invention.

Claims (6)

A detector that detects at least one of the energy, vibration, sound, heat, light, and power generated during processing during processing.
A feature amount extraction unit that extracts a first feature amount and a second feature amount from the detected value detected by the detection unit, and a feature amount extraction unit.
The point where the detected value is plotted on a plane with the numerical value related to the first feature amount as the first axis and the numerical value related to the second feature amount as the second axis, and the possibility that an abnormality may occur during the processing. A machine tool comprising at least two borders arranged in contour lines to represent, and a display for displaying. The machine tool according to claim 1, further comprising a correction unit for correcting the shape of the boundary line. The feature amount extraction unit includes a frequency decomposition unit that frequency-decomposes the detected value, a normalization unit that normalizes the data after frequency decomposition, and a dimension compression unit that compresses the dimension of the normalized data. The machine tool according to claim 1 or 2. Equipped with a holding part for holding tools for machining workpieces
The machine tool according to any one of claims 1 to 3, wherein the detection unit detects the detected value when the tool machins the work. The machine tool according to any one of claims 1 to 4, wherein the display unit displays information related to the plotted points when the plotted points are selected. Detected from a machine tool having a detection unit that detects a detection value of at least one of energy, vibration, sound, heat, light, and power generated during processing during processing. A display device that extracts the first feature amount and the second feature amount from the detected values and displays the machining status of the machine tool.
The point where the detected value is plotted on a plane with the numerical value related to the first feature amount as the first axis and the numerical value related to the second feature amount as the second axis, and the possibility that an abnormality may occur during the processing. A display device that displays at least two borders arranged in contour lines to represent.

Images