KR102373655B1 - Apparatus and method for diagnosing trouble of machine tool - Google Patents

Abstract

The present invention relates to a method for diagnosing a machine tool failure, comprising: a data collecting unit for collecting operation data of the machine tool; a PIP data generation unit for generating simplified PIP data using the data collected by the data collection unit using a PIP (Perceptually Important Points) method; a position adjusting unit for adjusting the position of the PIP data generated by the PIP data generating unit; and a diagnosis unit for diagnosing a failure of the machine tool by applying the PIP data whose position is adjusted by the position adjustment unit to a neural network.

Description

Machine tool failure diagnosis device and method

The present invention relates to a method for diagnosing machine tool failure, and more particularly, to a method for diagnosing machine tool failure by applying preprocessed operation data of the machine tool to a neural network.

A machine tool is a machine that processes a material to make a desired shape. In particular, in order to precisely process a material, such a machine tool is composed of a rotation system part that can move at a desired speed and a feed system part that can be transferred to an accurate position.

In order to improve the machining productivity and quality of these machine tools, machine tools are in the trend of high speed and high precision. As a technology for intelligence, a monitoring technology that can monitor the process status of a machine tool in real time is required, and a technology that determines the tool life of the machine tool is required.

In order to determine the life of a tool, it is necessary to measure the state of the tool, and these measuring methods can be broadly classified into two types: a direct method and an indirect method. The direct method is a technique for directly measuring tool wear using actual measuring equipment, and includes laser and ultrasonic waves. This technique has inefficiencies in terms of time and productivity, as it has to stop the production process and measure statically.

On the other hand, the indirect method is preferred because it can be measured in a dynamic state. Representatively, the measurement of acoustic emission (AE), the change in cutting force, the change in the load of the main shaft, the change using a vibrometer, and the measurement of illuminance are used. there is.

Tools that process materials in machine tools are factors that determine the quality of manufactured products, and tool breakage and tool wear are cited as the main causes of machine tool failure. It may cause a defect in the work piece. Tool breakage refers to a phenomenon in which a tool becomes unusable and causes fatal damage to workpieces and machine tools in the process of breaking. Tool wear is unavoidable because the tool is worn out due to friction in the process of producing the workpiece. The lifespan should be checked periodically.

The method of checking the current tool life is that workers in the field periodically stop the machine tools in the factory, check the dimensions of the workpiece, and in case of defects, directly check each tool to check the remaining life experience. Correct the position, if not, replace the tool. However, this method lacks accuracy and cannot accurately determine the remaining life.

Therefore, in the field, the maximum number of production possible for each tool is set at about 90%, and when it is reached, replacement is carried out. Such replacement and stopping of the machine tool to check the life of the tool reduces productivity, and the tool life is not used 100% and is replaced in advance to consume the tool unnecessarily.

Republic of Korea Patent Publication No. 10-1936283 (registered on January 02, 2019) Republic of Korea Patent Publication No. 10-2018-0048218 (published on May 10, 2018)

The present invention is to improve the problems in the prior art described above, and performs neural network learning using operation data of a machine tool, and provides an apparatus and a diagnostic method for diagnosing a failure related to tool wear of a machine tool using the neural network. Its purpose is to provide

In an apparatus for diagnosing a machine tool failure in the present invention for solving the above-described problem, the apparatus comprising: a data collecting unit for collecting operation data of the machine tool; a PIP data generation unit for generating simplified PIP data using the data collected by the data collection unit using a PIP (Perceptually Important Points) method; a position adjusting unit for adjusting the position of the PIP data generated by the PIP data generating unit; and a diagnosis unit for diagnosing a failure of the machine tool by applying the PIP data whose position is adjusted by the position adjustment unit to a neural network.

In addition, the data collection unit, a data classification unit for classifying the collected operation data into normal data and failure data; and a data generation unit that generates virtual failure data similar to the failure data.

In addition, the neural network is composed of LSTM (Long Short Term Memory)-RNN (Recurrent Neural Networks), and using the normal data and failure data classified by the data classification unit and the virtual failure data generated by the data generation unit, the Neural networks can be trained.

In addition, the PIP data generating unit may simplify using the PIP-ED method for diagnosing a point having the greatest distance between two adjacent points.

In addition, the position adjustment unit stores the x-axis data by finding the start point, the highest point, and the end point of each soaring part in the PIP data, and sets the target point based on the average of the highest point of each soaring part, from the starting point The position can be adjusted by moving the data up to the final point by an absolute value obtained by subtracting the target point from the highest point.

In addition, it is possible to determine the final failure only when the number of times determined to be a failure by performing the operation of the failure diagnosis apparatus of the machine tool is greater than or equal to a predetermined number of times.

A machine tool failure diagnosis method according to another embodiment of the present invention comprises the steps of: collecting operation data of the machine tool; generating simplified PIP data using the collected motion data in a PIP (Perceptually Important Points) method; adjusting the position of the PIP data; and diagnosing a failure of the machine tool by applying the adjusted PIP data to a neural network.

In addition, the step of collecting the operation data of the machine tool may include: classifying the collected operation data into normal data and failure data; and generating virtual failure data similar to the failure data.

In addition, the neural network is composed of LSTM (Long Short Term Memory)-RNN (Recurrent Neural Networks), and may further include the step of performing learning of the neural network using the normal data, failure data, and virtual failure data. .

In addition, the PIP method may be simplified by using the PIP-ED method for diagnosing a point having the greatest distance between two adjacent points.

In addition, the step of adjusting the position of the data may include: locating the starting point, the highest point, and the ending point of each soaring part in the PIP data and storing the x-axis data; setting a target point based on the average of the highest points of each of the soaring parts; and adjusting the position by moving the data from the starting point to the final point by an absolute value obtained by subtracting the target point from the highest point.

In addition, it is possible to determine the final failure only when the number of times determined to be a failure by performing the method for diagnosing the failure of the machine tool a plurality of times is greater than or equal to a predetermined number of times.

The method for diagnosing a machine tool failure according to the present invention as described above uses a neural network to diagnose a failure after performing pre-processing to remove unnecessary data such as noise, and finally determines the failure using the movement output frequency, resulting in a failure with high accuracy and reliability. can be judged

In addition, since the failure can be diagnosed using the real-time operation signal of the machine tool, there is no need to stop the operation of the machine tool for failure diagnosis, and the operation cost of the machine tool can be reduced because no manpower is required for failure diagnosis. can

In addition, it is possible to prevent fatal damage to the workpiece and machine by identifying tool damage such as tool breakage and tool wear of the machine tool, use the tool life to the maximum, and identify damage before fatal damage occurs. It has the effect of reducing the machine downtime by replacing it.

In addition, by preventing fatal damage to machine tools, it is possible to increase production, quality, and work efficiency of the factory, and by monitoring machine tool damage, it is possible to make a great contribution to factory automation. there is.

1 is a block diagram showing the structure of a failure diagnosis apparatus for a machine tool according to an embodiment of the present invention.
2A and 2B are diagrams illustrating a feature map of operation data for a failure diagnosis apparatus of a machine tool according to an embodiment of the present invention.
3 is a diagram schematically illustrating a calculation method of PIP methods used in an embodiment of the present invention.
4 is a view for explaining a position adjustment algorithm of the apparatus for diagnosing a machine tool failure according to an embodiment of the present invention.
5A to 5E are diagrams for explaining a method of operating a neural network of an apparatus for diagnosing a failure of a machine tool according to an embodiment of the present invention.
6 is a flowchart illustrating a neural network learning method for diagnosing a failure of a machine tool according to an embodiment of the present invention.
7 is a flowchart illustrating a fault diagnosis method for fault diagnosis of a machine tool according to an embodiment of the present invention.

Since the present invention can apply various transformations and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the specific embodiments of the present invention, and it should be understood to include all modifications, equivalents and substitutes included in the spirit and scope of the present invention. In describing the present invention, if it is determined that a detailed description of a related known technology may obscure the gist of the present invention, the detailed description thereof will be omitted.

Terms such as first, second, etc. may be used to describe various elements, but the elements should not be limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

Terms used in the present invention are only used to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as “comprise” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It should be understood that this does not preclude the existence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Spatially relative terms such as below, beneath, lower, above, etc. will be used to easily describe the correlation between one component and other components, as shown in the drawings. can Spatially relative terms should be understood as terms including different directions of components when used in addition to the directions shown in the drawings. For example, when a component shown in the drawing is turned over, a component described as below (below, beneath) other components may be placed above (above, upper) other components. Accordingly, the example term below may include both downward and upward directions. Components may also be oriented in other orientations, and thus spatially relative terms may be interpreted according to orientation.

As used in the present invention, the expression indicating a part such as “part” or “part” means a device in which the component may include a specific function, software that may include a specific function, or a device that may include a specific function. And it means that it can represent a combination of software, but it is not necessarily limited to the expressed function, which is provided to help a more general understanding of the present invention, and a person with ordinary knowledge in the field to which the present invention belongs If it is, various modifications and variations are possible from these base materials.

Therefore, the spirit of the present invention should not be limited to the described embodiments, and not only the claims to be described later but also all of the claims and equivalents or equivalent modifications will be said to belong to the scope of the spirit of the present invention. .

Hereinafter, a method for diagnosing a machine tool failure according to a preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings.

Recently, a high-precision machine tool is attached to a CNC (computer numerical control) to set the position of the tool with respect to the workpiece. CNC is a system for controlling machining by machine tools such as lathes and cutting machines using a computer. In general, when the numerical size of a workpiece designed with a program such as AutoCAD is transmitted to the machine tool, the CNC of the machine tool can calculate the numerical size of the workpiece. It can be collectively referred to as a system that cuts and processes the tool by setting the position of the tool accordingly. Most of these CNCs include built-in Ethernet, and such Ethernet can communicate with an external system such as a computer using a power line or a communication line.

In an embodiment of the present invention, communication may be performed through such a built-in Ethernet, but if there is no built-in Ethernet, a communication unit (not shown) may be additionally installed to communicate with an external system.

In an embodiment of the present invention, the operation data refers to the spindle torque data of the machine tool, but is not limited thereto, and the error rate of the workpiece that can be collected while the machine tool operates, the operation signal of the motor, the operation time of the machine tool, and the noise All status information generated by machine tools such as , vibration, etc. can be used.

In one embodiment of the present invention, the failure is typically described by setting the wear of the machine tool of the machine tool as a representative embodiment, but is not limited thereto, and may mean all parts that require fault diagnosis of the machine tool.

1 is a block diagram showing the structure of a failure diagnosis apparatus for a machine tool according to an embodiment of the present invention.

The failure diagnosis apparatus in one embodiment of the present invention is configured to include a data collection unit 100, a preprocessor 200, a diagnosis unit 300, and a determination unit 400 to diagnose a failure of a machine tool in real time. can be

The data collection unit 100 may collect data in real time through the CNC built-in Ethernet of the machine tool. The data collected in this way is packet data, and the packet data may include state information of the machine tool and torque data of the main shaft motor.

The status information is generally transmitted in units of 8 bits (1 byte), for example, “01” is an emergency stop, “09” is waiting after processing the workpiece, and “0B or 6B” is the processing of the workpiece. It can be expressed as a medium signal or the like.

In addition, the torque data of the main shaft motor may be composed of 16 bits (2 bytes), and expressed as a decimal number may be expressed as a number from 0 to 65,535.

In an embodiment of the present invention, talk data among the packet data may be collected as motion data.

When the data collection unit 100 collects motion data, the data classification unit 110 classifies the data, and based on the classified data, the data generation unit 120 may generate learning data for neural network learning.

The data classification unit 110 may classify the torque data from packet data obtained from Ethernet in the CNC.

The torque data classified by the data classifying unit 110 may be largely classified into two types of torque data, and the two types are normal torque data and faulty torque data.

Such normal torque data and fault torque data may also be classified by the data classifying unit 110 .

2 is a view showing a characteristic map of operation data for a failure diagnosis apparatus of a machine tool according to an embodiment of the present invention. Specifically, FIG. 2a is a diagram showing actual normal torque data classified by the data classification unit 110 and actual failure It is torque data, and FIG. 2B shows virtual fault torque data generated by the data generator 120 in addition to the actual normal torque data and the actual fault torque data.

Specifically, referring to FIG. 2A , the data classifying unit 110 classifies normal torque data and fault torque data with respect to the collected data, blue O is actual normal torque data, and red X represents actual fault torque data. there is. As can be seen in FIG. 2A , the number of normal torque data is overwhelmingly large, and accordingly, the number of failure torque data as a sample for performing neural network learning may be relatively small.

The data generator 120 may generate virtual fault torque data in order to resolve such data imbalance. Referring to FIG. 2B , the actual normal torque data is indicated by a blue O, the actual failure torque data is indicated by a black X, and the virtual failure torque data is indicated by a green ◇. The virtual failure torque data is data having similar characteristics to the actual failure torque data, and may be used as data for neural network learning.

That is, in the process of performing neural network learning of the present invention, when data is transmitted from the machine tool, the data collection unit 100 classifies the data in the data classification unit 110 into actual normal torque data and actual failure torque data, and generates data After generating the virtual failure torque data using the actual failure torque data in the unit 120, learning is performed using the real normal torque data, the actual failure torque data, and the virtual failure torque data, and the machine tool using the learned neural network In the process of diagnosing the failure of the machine tool, when data is transmitted from the machine tool, the data collection unit 100 collects data, and it can be transferred to the preprocessor 200 immediately.

Alternatively, if necessary, in the process of diagnosing the failure of the machine tool, the actual normal torque data and the actual failure torque data may be first classified using the data classification unit 110 and then confirmed using a neural network. form can be used.

The data classifying unit 110 and the data generating unit 120 of the data collecting unit 100 may classify and generate data based on preset standards, but when performing these operations, accurate data classification and data generation are performed. Reliability problems may occur to perform.

In an embodiment of the present invention, the data classifying unit 110 and the data generating unit 120 may be configured as a generative adversarial network (GAN) for accurate data classification and generation.

GAN is a neural network largely composed of a generator and a comparison part. The generator generates similar data to the input data, and the comparison part compares the input data with the generated data to generate aborted data. is a neural network

The data classification unit 110 classifies the operation data of the machine tool into normal data and failure data by using the GAN learned in advance, and the data generation unit 120 can generate similar data by using the generation unit of the GAN. .

The data collected by the data collection unit 100 is pre-processed by the pre-processing unit 200 to diagnose a machine tool failure in the diagnosis unit 300 through the PIP data generation unit 210 and the position adjustment unit 220 . can

Data preprocessing is to increase operational efficiency by processing noise-containing data, incomplete data, and contradictory data. A method of consolidating data into a single data that exists, a method of improving accuracy and efficiency by logging, averaging, or binning data through data transformation, or a method of reducing the size of data through processes such as aggregation, deduplication, and clustering etc. can be used.

Motion data used in an embodiment of the present invention is time-series data and has high dimensionality and continuity, and thus contains a lot of noise, which may increase the amount of calculation. In order to increase the processing efficiency of such time series data, a PIP (Perceptually Important Points) method can be used to reduce the size of the data while maintaining important characteristic points of the data.

The PIP data generation unit 210 reduces the size of the x-axis and the y-axis of the time series operation data of the machine tool collected by the data collection unit 100 by using the PIP method, and reduces the size of the data by simplifying the form. However, it is possible to maintain important characteristics of the data.

The PIP method may include three types of PIP-VD (Vertical Distance), PIP-PD (Perpendicular Distance), and PIP-ED (Euclidean Distance) methods, and will be described in detail with reference to FIG. 3 below. .

3 (a) relates to the PIP-VD method, which is a method for diagnosing the point that is vertically the furthest from a straight line connecting two adjacent PIPs and a straight line drawn in the y-axis direction. It can be obtained using Equation 1.

In Equation 1, y _k may be a y-axis coordinate, and x _k may mean an x-axis coordinate.

3 (b) relates to a PIP-PD method, which is a method for diagnosing a point that is vertically furthest from a straight line connecting two adjacent PIPs, and can be obtained using Equation 2 below.

In Equation 2, y _k may be a y-axis coordinate, and x _k may mean an x-axis coordinate.

3(c) relates to a PIP-ED method, which is a method for diagnosing a point having the longest distance between two adjacent points, which may be obtained using Equation 3 below.

In Equation 3, y _k may be a y-axis coordinate, and x _k may mean an x-axis coordinate.

In an embodiment of the present invention, preprocessing for simplifying operation data is performed using the PIP-ED method, which is the method (c) of FIG. 3 , but the present invention is not limited thereto, and PIP-VD and PIP-PD methods may be used. .

The data pre-processed by the PIP data generation unit 210 preserves the main characteristics by simplifying it by using the PIP method, but the actual data and its shape and location are changed. Since an error may occur in the process of diagnosing a machine tool failure due to such a position change, the position adjustment unit 220 should adjust the position to improve accuracy.

4 is a diagram for adjusting the position of the PIP data generated by the PIP data generation unit 210 through the position adjustment algorithm in the position adjustment unit 220 according to an embodiment of the present invention.

,

, ...,

), 초록색으로 표시된 최고점(

,

, ...,

), 파란색으로 표시된 종료점(

,

, ...,

)을 찾은 뒤에 시작점, 최고점, 종료점 각각의 x축 데이터를 저장할 수 있다. The method of adjusting the position of PIP data is described with reference to FIG. 4, the starting point (

,

, ...,

), the highest point in green (

,

, ...,

), the end point shown in blue (

,

, ...,

), you can save the x-axis data of each starting point, peak, and ending point.

,

, ...,

) 평균을 기준으로 타겟점(

,

, ...,

)을 설정할 수 있다. 타겟점이 설정되면, (

)부터 종료점(

)까지의 데이터를 최고점과 타겟점 간의 차이의 절댓값(

)만큼 이동하여 위치를 조절할 수 있다.The highest point of each soaring part of the stored data (

,

, ...,

) based on the average of the target point (

,

, ...,

) can be set. When the target point is set, (

) to the end point (

) to the absolute value of the difference between the highest point and the target point (

) to adjust the position.

That is, if the difference value between the highest point and the target point is positive, it moves from the start point, and if the difference value between the highest point and the target point is negative, it moves from the end point to remove the data overlapping situation. The position of the PIP data can be adjusted by repeating up to the second value.

The operation data that has been pre-processed in the pre-processing unit 200 may diagnose a machine tool failure in the diagnosis unit 300 .

The diagnosis unit 300 may be configured as a Recurrent Neural Network (hereinafter, LSTM-RNN) using a Long Short Term Memory (LSTM) memory cell.

An artificial neural network is a deep learning model for learning data that changes over time, such as time-series data. Unlike general neural networks that perform calculations in the hidden layer and output in the output layer, the hidden values are stored in the memory located inside the neural network and then used as input data for the next order to learn We can talk about the neural network used for Due to these structural features, the cyclic neural network can be effectively utilized when data including time information, such as handwriting, voice signals, and sensor data, exist.

A recurrent neural network (hereinafter, RNN), which is one type of such artificial neural networks, can typically utilize a method using an RNN basic memory cell (hereinafter, an RNN memory) or using an LSTM.

The RNN memory can generate an output by calculating the input of the base time and the output of the previous time by using the tanh function, which is a hyperbolic tangent. However, in the RNN memory, there may be a problem in that the vanishing gradient problem occurs, in which the main information cannot be remembered as it passes through several time steps in the learning process, and the progress of learning may be stopped. In other words, if the amount of change is small due to the gradient loss problem, the gradient becomes almost 0, and as the training progresses, the main information cannot be remembered. There may be problems with losing.

In order to solve the above RNN memory problem, a method of using an LSTM as a memory cell of an RNN is emerging, and an embodiment of the present invention can also be operated by using an RNN using an LSTM and an LSTM-RNN.

Hereinafter, an operation method of the LSTM-RNN will be described with reference to FIGS. 5A to 5E .

5A is an overall view of a memory cell used in an LSTM-RNN, and the LSTM-RNN may include four gates.

,

와 바이어스의 가중 합에 시그모이드 함수(Sigmoid function)를 거친 형태로 계산을 통해

를 구하는 동작을 수행하여 해당 정보를 기억할지 잊을지 결정할 수 있다.5B is a diagram illustrating an operation of a Forget Gate, which is a first gate among four gates, and the Forget Gate may determine whether to store a cell state. Forget Gate is as follows using Equation 4 below

,

Through calculation in the form of a sigmoid function on the weighted sum of and bias

It is possible to decide whether to remember or forget the information by performing an operation to obtain

In Equation 4, ft may be an output of the Forget gate, σ may be a sigmoid function, Wf may be a weight, ht-1 may be a previous short-term state, xt may be a current input, and bf may be a bias of the Forget gate.

)을 결정할 수 있으며, 하기의 수학식 6을 이용하여 업데이트할 값(

)을 셀에 갱신할지 결정할 수 있다.FIG. 5C is a diagram illustrating an operation of an input gate, which is a second gate among four gates, and the input gate may determine whether to store new information in a cell state. The input gate uses the following Equation 5 to pass the tanh function, which is a hyperbolic tangent, to the value to be updated in the cell (

) can be determined, and the value to be updated using Equation 6 below (

) to update the cell.

는 장기 상태 후보 값이고, W_C는 가중치, h_t-1은 이전의 단기 상태, x_t는 현재 입력, b_C는 편향일 수 있다. In Equation 5 above,

may be a long-term state candidate value, W _C may be a weight, h _t-1 may be a previous short-term state, x _t may be a current input, and b _C may be a bias.

In Equation 6, i _t may be an output of the update gate, σ may be a sigmoid function, h _t-1 may be a previous short-term state, x _t may be a current input, and b _i may be a bias of the input gate.

를 더한 값을 업데이트할 값(C_t)으로 갱신하여 셀에 저장될 수 있다.5D is a view showing the operation of the Update Gate, which is the third gate among the four gates. The Update Gate updates the value to be updated at the previous time (C _t-1 ) to the value to be updated at the current time (C _t ). can do. After the update gate is multiplied by f _t by C _t-1 using Equation 7 below,

It can be stored in the cell by updating the value plus the value to be updated (C _t ).

는 장기 상태 후보 값일 수 있다.In Equation 7, C _t is the long-term state value, f _t is the output of the Forget gate, C _t-1 is the long-term state value in the previous time period, i _t is the bias of the Update gate,

may be a long-term state candidate value.

5E is a diagram illustrating an operation of an Output Gate, which is a fourth gate among four gates, and the Output Gate may be a gate for outputting information. The Output Gate has a value between -1 and 1 using Equation 9 below after going through a sigmoid function to determine which part to output in the cell state using Equation 8 below. Information can be output through the tanh function, which is a hyperbolic tangent to the cell state.

In Equation 8, O _t is the output of the output gate, σ is a sigmoid function, W _o is a weight, h _t-1 is a previous short-term state, x _t is a current input, and b _o is an output gate bias. .

In Equation 9, h _t-1 may be a previous short-term state, O _t may be an output of an output gate, and C _t may be a long-term state value to be updated.

The diagnostic unit 300 according to an embodiment of the present invention may be configured as an LSTM-RNN including the operation process as described above, and the LSTM-RNN includes 150 input layers and 100 LSTM memory cells, including a hidden layer and an output layer. can be composed of The input layer, the hidden layer, and the output layer are not fixed values and can be flexibly changed or set according to the size, period, and measurement signal interval of the operation data to be processed. However, all types of processors that can improve performance through various types of learning such as artificial neural networks, convolutional neural networks, machine learning, and deep learning can be used.

In an embodiment of the present invention, since the diagnosis unit 300 is composed of an LSTM-RNN, learning of the neural network may be required.

The RNN model used in an embodiment of the present invention may be constructed using a many-to-one structure and an LSTM memory cell structure.

In the many-to-one structure, RNN is performed by tokenizing time series data, and when the output of the last token is generated, the loss between the output and the correct answer (normal value) is calculated and then the error backpropagation method is used. It is a structure to find the correct answer by minimizing the loss function by proceeding with learning.

The LSTM-RNN used in an embodiment of the present invention performs PIP-ED processing on real torque data and virtual torque data to generate PIP data from which data is reduced and noise is removed, and the position of PIP data is adjusted. The data can be used as input data for the neural network. When the positioning data is input to the neural network through the input layer, the neural network operation is performed in the hidden layer through the LSTM memory cell and the loss is calculated. Branches can be certified as Fully-Connected. Fully-Connected connects the outputs from all nodes and passes them to the sotfmax function located in the output layer, then the softmax function receives the output value and outputs it as a probability vector to calculate the output loss. The output loss calculated through the Softmax function can be learned through the loss difference between the input and the output in the classifier.

In an embodiment of the present invention, the data collection unit 100 uses the virtual failure torque data generated by the data generation unit 120 together with the actual normal torque data and the actual failure torque data classified by the data classification unit 110 . Learning of LSTM-RNN can be performed.

In addition, when the learning of the diagnosis unit 300 is completed, the failure of the machine tool may be diagnosed by using the data preprocessed by the preprocessing unit 200 .

At this time, the machine tool operation data collected by the data collection unit 100 is directly pre-processed by the preprocessing unit 200 to diagnose a machine tool failure in the diagnosis unit 300 or the machine tool collected by the data collection unit 100 . After first classifying the machine operation data by the data classification unit 110 , only the data suspected of a failure may be pre-processed in the pre-processing unit 200 to diagnose the operation of the machine tool in the diagnosis unit 300 .

Even if a failure is diagnosed using an RNN, an error may occur in classifying normal data as failure data or classifying failure data as normal data because the diagnosis is not performed with 100% accuracy.

A failure occurred due to such an error, but excessively it causes a major failure of the machine tool, or if it is judged as a failure during normal operation and stops the operation, it may cause a setback in production and economic loss. It is important that the method has 100% accuracy as much as possible.

In one embodiment of the present invention, in order to accurately diagnose the failure of the machine tool with the highest possible probability, the determination unit 400 determines whether the failure of the machine tool diagnosed by the diagnosis unit 300 is normally diagnosed using the movement output frequency. can

The movement output frequency may refer to a method of outputting a result value when the number of determination criteria Ns is satisfied within the determination reference number with respect to the recently operated determination reference number N.

In an embodiment of the present invention, the movement output frequency is determined by the determination unit 400 as a final failure when Ns or more of the N normal or failure diagnosis results diagnosed by the diagnosis unit 300 are diagnosed as a failure, and a failure is provided to the user. It may be a method of outputting a signal.

For example, referring to Table 1 below, if N is set to 5 and Ns is set to 3, and three or more of the five diagnostic result data recently diagnosed as a failure, the final result is regarded as a failure. It is about the number of cases to judge.

In an embodiment of the present invention, N is set to 5 and Ns is set to 3, but this is not fixed, and the operation time of the machine tool, the frequency of recent failures, the precision of the work, the probability of occurrence of failure diagnosis, and the user You can change the settings based on your experience.

이고, 고장으로 분류될 확률은

이며, 상기 표1의 고장상태일 확률을 이용하여 계산하면 오류 발생 확률을 예측할 수 있다. 예를 들어, 상기 표 1의 8번 상태인 정상, 정상, 고장, 고장, 고장으로 판단되어 최종상태가 고장상태로 판단되었으며, 8번 상태의 오류율은

로 계산되어 최종적으로 0.081, 8.1%의 오류 발생확률을 포함할 수 있다. Referring to Table 1 above, when the accuracy of the RNN is 90%, the probability of being classified as normal is

and the probability of being classified as a failure is

, and the error occurrence probability can be predicted by calculating using the probability of a failure state in Table 1. For example, the 8th state of Table 1 was determined as normal, normal, faulty, faulty, and failed, and the final state was determined to be a faulty state, and the error rate of the 8th state was

It is calculated as , and can finally include the probability of error occurrence of 0.081 and 8.1%.

Explaining again based on the above example, when the RNN accuracy of the diagnosis unit 300 is 90%, that is, the error occurrence probability is 10%, the determination unit 400 uses the moving output frequency to reduce the error occurrence probability to 8.1 It can be confirmed that it can be lowered to %, and as the accuracy of the RNN increases, the number of judgment criteria of the moving output frequency increases, and as the number of samples increases, the error occurrence probability decreases in inverse proportion.

That is, the present invention diagnoses a machine tool failure using a neural network, and can more accurately diagnose a machine tool failure by combining the results of the neural network decision to make a final decision in order to increase reliability.

Finally, when the failure or normality of the machine tool is determined, the determination result may be stored in the form of notifying the administrator of the result using a display or a mobile terminal or storing the result in a server (not shown). In addition, when it is determined that a failure is detected, a device such as a warning light or a speaker is attached to the machine tool to immediately notify the surroundings with sound, light, etc.

6 is a flowchart illustrating a neural network learning method for diagnosing a failure of a machine tool according to an embodiment of the present invention.

Data can be collected in real time through the CNC built-in Ethernet of the machine tool (S110).

The collected data may include state information of the machine tool and torque data of the main shaft motor as packet data. In an embodiment of the present invention, a method for diagnosing a failure using only the torque data of the main shaft motor is described. Without limitation, it is possible to diagnose a failure by collecting various data related to the machine tool.

The collected data may be classified into actual normal data and actual failure data (S120).

In order to classify the collected data into actual normal data and actual failure data, a generative adversarial neural network (GAN) can be used to classify the collected data.

Since the number of actual failure data is relatively small in contrast to the actual normal data among the collected data, there may be a limit to performing neural network learning, and for this purpose, virtual failure data may be generated (S130).

The virtual failure data can be generated by using the generator in the generative adversarial neural network (GAN), but is not limited thereto, and similar data such as machine learning, deep learning, convolutional neural network, artificial neural network, and SVM can be generated. If anything, you can use it.

The actual normal data, the actual failure data, and the virtual failure data may be pre-processed to remove noise and simplify the PIP method (S140).

In an embodiment of the present invention, simplification is performed by using the PIP-ED method, but PIP-VD and PIP-PD methods, which are PIP methods, may be used.

For a detailed PIP method, refer to the description of FIG. 3, and a detailed description thereof will be omitted.

The PIP data simplified through the PIP method can be moved to the left and right at a certain interval compared to the original position during the simplification process, and when using these data directly, accurate learning cannot be performed due to an error, so the positioning algorithm can be used to adjust the position of the PIP data (S150).

The positioning algorithm sets the target point based on the average of the highest points after finding the starting point, the highest point, and the rising end point of the rising part in the PIP data, and moves the starting point and the ending point in the x-axis direction by the difference between the target point and the highest point. way to control it.

The position-adjusted PIP data may be applied to a regression neural network composed of LSTM memory cells to perform learning (S160).

7 is a flowchart illustrating a diagnosis method for diagnosing a failure of a machine tool according to an embodiment of the present invention.

Data can be collected in real time through the CNC built-in Ethernet of the machine tool (S210).

The collected data may include state information of the machine tool and torque data of the main shaft motor as packet data. In an embodiment of the present invention, a method for diagnosing a failure using only the torque data of the main shaft motor is described. Without limitation, it is possible to diagnose a failure by collecting various data related to the machine tool.

The collected data may be pre-processed immediately or primary classification may be performed using the data classification unit 110 configured as a GAN, and an operation may be set according to the user's convenience.

The collected data may be subjected to pre-processing to reduce noise while simplifying the noise through the PIP method (S220).

In an embodiment of the present invention, simplification is performed by using the PIP-ED method, but PIP-VD and PIP-PD methods, which are PIP methods, may be used.

For a detailed PIP method, refer to the description of FIG. 3, and a detailed description thereof will be omitted.

The PIP data simplified through the PIP method can be moved to the left and right at a certain interval compared to the original position during the simplification process, and when using these data directly, accurate learning cannot be performed due to an error, so the positioning algorithm can be used to adjust the position of the PIP data (S230).

The positioning algorithm sets the target point based on the average of the highest points after finding the starting point, the highest point, and the rising end point of the rising part in the PIP data, and moves the starting point and the ending point in the x-axis direction by the difference between the target point and the highest point. way to control it.

The position-adjusted PIP data is applied to a regression neural network composed of LSTM memory cells to diagnose machine tool failure (S240).

Even if neural network learning is performed using a lot of sample data, the judgment of the neural network cannot have 100% accuracy, which may cause an error in classifying normal data as faulty data or classifying faulty data as normal data.

In order to increase the failure diagnosis accuracy, a final failure may be determined using the movement output frequency (S250).

The moving output frequency is a method of collecting the determination result of a certain number of times, calculating a failure or normal rate from the collected results, comparing it with a preset rate, and providing the final result. In an embodiment of the present invention, the latest N failure diagnosis Combining the results, it can be judged as a final failure if Ns or more failures are found among the N number of failure diagnosis results.

As described above, the present invention collects the operation data of the machine tool, pre-processes it by the PIP method, adjusts the position of the PIP data, diagnoses the failure using the LSTM-RNN, and determines the final failure using the movement output frequency. By judging, failures can be determined with high accuracy and reliability, and by collecting data in real time to determine failures, unnecessary maintenance time and waste of manpower for failure diagnosis can be prevented.

In addition, by accurately diagnosing a failure in real time, it is possible to immediately identify tool damage such as tool breakage and tool wear of the machine tool, thereby preventing fatal machine damage, and bringing economic benefits by maximizing the life of the tool.

So far, with respect to the present invention, the preferred embodiments have been looked at. Those of ordinary skill in the art to which the present invention pertains will understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments are to be considered in an illustrative rather than a restrictive sense. The scope of the present invention is indicated in the claims rather than the foregoing description, and all differences within the scope equivalent thereto should be construed as being included in the present invention.

100: data collection unit
110: data classification unit
120: data generation unit
200: preprocessor
210: PIP data generation unit
220: position adjustment unit
300: detection unit
400: decision part

Claims (12)

A data collection unit for collecting operation data of the machine tool;
a PIP data generation unit for generating simplified PIP data using the data collected by the data collection unit using a PIP (Perceptually Important Points) method;
a position adjusting unit for adjusting the position of the PIP data generated by the PIP data generating unit; and
a diagnosis unit for diagnosing a failure of the machine tool by applying the PIP data whose position is adjusted by the position adjustment unit to a neural network;
A fault diagnosis device comprising a. According to claim 1,
The data collection unit,
a data classification unit for classifying the collected operation data into normal data and failure data; and
a data generation unit for generating virtual failure data similar to the failure data;
A fault diagnosis device comprising a. 3. The method of claim 2,
The neural network is
Consists of Long Short Term Memory (LSTM)-RNN (Recurrent Neural Networks),
The failure diagnosis apparatus according to claim 1, wherein the neural network is learned by using the normal data and failure data classified by the data classification unit and the virtual failure data generated by the data generation unit. According to claim 1,
The PIP data generation unit,
A failure diagnosis device, characterized in that it is simplified using the PIP-ED method for diagnosing the point that is the furthest between two adjacent points. According to claim 1,
The positioning unit,
Find the start point, the highest point, and the end point of each soaring part in the PIP data, store the x-axis data, set the target point based on the average of the highest point of each soaring part, and the data from the starting point to the highest point A failure diagnosis apparatus, characterized in that the position is adjusted by moving by an absolute value obtained by subtracting the target point from the highest point. According to claim 1,
A failure diagnosis apparatus, characterized in that the final failure is determined only when the number of times determined as failures by performing the operation of the failure diagnosis apparatus of the machine tool is greater than or equal to a predetermined number of times. In the method for diagnosing a machine tool failure,
collecting operation data of the machine tool;
generating simplified PIP data using the collected motion data in a PIP (Perceptually Important Points) method;
adjusting the position of the PIP data; and
diagnosing a failure of the machine tool by applying the adjusted PIP data to a neural network;
A fault diagnosis method comprising a. 8. The method of claim 7,
The step of collecting operation data of the machine tool comprises:
classifying the collected operation data into normal data and failure data; and
generating virtual failure data similar to the failure data;
A fault diagnosis method comprising a. 9. The method of claim 8,
The neural network is
Consists of Long Short Term Memory (LSTM)-RNN (Recurrent Neural Networks),
The method for diagnosing failure, characterized in that it further comprises the step of performing learning of the neural network using the normal data, the failure data, and the virtual failure data. 8. The method of claim 7,
The PIP method is
A fault diagnosis method characterized in that it is simplified using the PIP-ED method for diagnosing the point that is the furthest between two adjacent points. 8. The method of claim 7,
Adjusting the location of the data includes:
finding the start point, the highest point, and the end point of each soaring part in the PIP data and storing the x-axis data;
setting a target point based on the average of the highest points of each of the soaring parts; and
adjusting the position by moving the data from the starting point to the highest point by an absolute value obtained by subtracting the target point from the highest point;
A fault diagnosis method comprising a. 8. The method of claim 7,
A failure diagnosis method, characterized in that the final failure is determined only when the number of times determined to be a failure by performing the method for diagnosing the failure of the machine tool a plurality of times is greater than or equal to a predetermined number of times.

Images