JP6148316B2 - Machine learning method and machine learning device for learning failure conditions, and failure prediction device and failure prediction system provided with the machine learning device - Google Patents

Description

  The present invention relates to a machine learning method and a machine learning device for learning a failure condition, and a failure prediction device and a failure prediction system including the machine learning device.

  In an industrial machine, it may be required to detect abnormality of a component in advance in order to improve the yield or prevent the occurrence of a serious accident. For example, a method of comparing an output value of a sensor with a predetermined threshold value and detecting an abnormality based on the result is known. Here, the term “industrial machine” means not only an industrial robot and a machine controlled by a computer numerical control (CNC) device, but also a service robot and a machine including various mechanical devices. And

  Patent Document 1 discloses a failure prediction diagnosis method for predicting a failure of a robot by comparing a reference operation pattern of a normal robot and an operation pattern of a robot in operation.

  Japanese Patent Laid-Open No. 2004-26883 discloses a robot by comparing a difference between a load-side power based on an actual operation state of a drive shaft and a drive-side power based on an operation command to the drive shaft with a determination value. A failure prediction method for evaluating the presence / absence and deterioration level of a mechanism unit is disclosed.

JP 63-123105 A Japanese Patent Application Laid-Open No. 10-039908

  However, factors that lead to failures are becoming more complex as industrial machines become more complex or sophisticated. Therefore, the conventional failure prediction method executed according to a certain standard may not be applicable to the actual situation or may not be accurate. Therefore, there is a demand for a failure prediction device that enables accurate failure prediction according to the situation.

According to the first invention of the present application, there is provided a machine learning device that learns a condition associated with a failure of an industrial machine, and controls output data of a sensor that detects a state of the industrial machine or an ambient environment, and the industrial machine. A state observation unit for observing an internal data of the control software and a state variable including at least one of the output data or calculation data obtained based on the internal data while the industrial machine is operating or stationary; and the industry Learning a condition associated with a failure of the industrial machine according to a determination data acquisition unit that acquires determination data for determining a degree of machine failure and a training data set created based on a combination of the state variable and the determination data Bei example a learning section, a for said learning section, the determination data acquisition unit has acquired the decision data indicating a failure of the industrial machinery To come, the determination data contained in the training data set, machine learning unit for updating the condition by weighting depending from when a failure occurs on the length of time going back until obtaining the judgment data is provided .
According to a second invention of the present application, in the machine learning device according to the first invention, the learning unit is configured to learn the condition according to the training data set created for a plurality of industrial machines. Is done.
According to the third invention of the present application, in the machine learning device according to the first or second invention, the learning unit learns a normal state only in a certain period, and thereafter, the failure by the determination data acquisition unit Ru is configured to detect an occurrence.
According to the fourth aspect of the invention, with a machine learning device according to any one of the first to third aspects, a failure prediction device for predicting the failure of the industrial machine, wherein the learning unit is the training Based on the learning result according to the data set, the failure prediction unit further includes a failure information output unit that outputs failure information indicating the presence or absence of the failure of the industrial machine in response to the current input of the state variable. An apparatus is provided.
According to the fifth aspect of the invention, the failure prediction apparatus according to the fourth invention, the learning section in accordance with additional training data set that is created based the on the combination of the current state variables and the decision data, It is configured to relearn the condition.
According to the sixth aspect of the invention, the failure prediction apparatus according to the fourth or fifth invention, the machine learning system is connected to the industrial machine via a network, said state observing unit, through the network Configured to obtain the current state variable.
According to the seventh invention of the present application, in the failure prediction device according to the sixth invention, the machine learning device exists on a cloud server.
According to the eighth aspect of the invention, the failure prediction apparatus according to any one of the invention from the 4 th 6 th, the machine learning system is built in the control apparatus for controlling the industrial machine.
According to the ninth aspect of the invention, the failure prediction apparatus according to any one of the invention from the 4 th 8 th, learning result by the machine learning device is shared by a plurality of said industrial machine.
According to the tenth aspect of the invention, a failure prediction apparatus according to any one of the invention from the 4 th ninth, a sensor for outputting the output data, and fault information notification unit that notifies the failure information to the operator, A failure prediction system is provided.
According to the eleventh invention of the present application, in the failure prediction system according to the tenth invention, the time when the failure information is notified to the operator by the failure information notification unit is the first from the time when the failure occurs. It is before the time determined in a predetermined period.
According to the 12th aspect of the invention, the failure prediction system according to the 11th invention, timing of the failure information in the failure information notifying unit is notified to the operator, failure first back from time to occur It is before the time determined in the predetermined period and after the time determined in the second predetermined period that is longer than the first predetermined period, going back from the time when the failure occurs.
According to a thirteenth aspect of the present invention, there is provided a machine learning method for learning a condition associated with a failure of an industrial machine, wherein the output data of a sensor for detecting the state of the industrial machine or the surrounding environment, and the industrial machine are controlled. A state variable including at least one of internal data of the control software and calculation data obtained based on the output data or the internal data is observed during operation or stationary of the industrial machine, and failure of the industrial machine is detected. Acquires determination data for which the degree has been determined, learns a condition associated with the failure of the industrial machine according to a training data set created based on the combination of the state variable and the determination data, and associates with the failure of the industrial machine Is included in the training data set when determination data representing failure of the industrial machine is acquired. It said decision data, machine learning method for updating the condition by weighting depending from when a failure occurs on the length of time going back until obtaining the judgment data is provided that.

  These and other objects, features and advantages of the present invention will become more apparent by referring to the detailed description of the exemplary embodiments of the present invention shown in the accompanying drawings.

  The machine learning device and the machine learning method according to the present invention learn a condition associated with a failure of an industrial machine according to a training data set created based on a combination of state variables and determination data. Since the failure condition is learned while the industrial machine is actually operated, an accurate failure condition according to the actual use situation is learned. In addition, according to the failure prediction apparatus and the failure prediction system according to the present invention, since the machine learning device capable of machine learning of the failure condition is provided, accurate failure prediction according to the actual use situation becomes possible.

FIG. 1 is a block diagram illustrating an example of a failure prediction system according to an embodiment. FIG. 2 is a flowchart illustrating an example of a learning process flow in the machine learning apparatus. FIG. 3 is a diagram illustrating a configuration example of a neural network. FIG. 4 is a diagram for explaining an example of a learning period in an unsupervised learning technique. FIG. 5 is a diagram for explaining an example of a recurrent neural network. FIG. 6 is a block diagram illustrating an example of a failure prediction system according to another embodiment. FIG. 7 is a diagram (part 1) for explaining an example of an index value indicating the degree of failure in the failure prediction system according to the embodiment. FIG. 8 is a diagram (No. 2) for explaining an example of the index value indicating the degree of failure in the failure prediction system according to the embodiment. FIG. 9 is a flowchart illustrating an example of a flow of failure prediction using the learning result.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of a machine learning method and a machine learning device according to the present invention, and a failure prediction device and a failure prediction system including the machine learning device will be described with reference to the accompanying drawings. The components of the illustrated embodiment are appropriately scaled to assist in understanding the present invention. The same reference numerals are used for the same or corresponding components.

  FIG. 1 is a block diagram illustrating an example of a failure prediction system according to an embodiment. The failure prediction system 1 can learn a condition (hereinafter, also referred to as “failure condition”) associated with an industrial machine failure using the machine learning device 5 having a machine learning function. Further, the failure prediction system 1 can create failure information according to the state of the industrial machine and its surrounding environment based on the result learned by the machine learning device 5.

  In this specification, “industrial machine” means various machines including an industrial robot, a service robot, and a machine controlled by a computer numerical control (CNC) device. Further, in the present specification, “failure of an industrial machine” includes a failure of a component of an industrial machine. That is, the “failure of the industrial machine” is not limited to a state where the intended function of the industrial machine cannot be executed, and includes, for example, a state where the normal operation cannot be temporarily or permanently reproduced.

  The “failure information” created by the failure prediction system 1 includes information indicating the presence or absence of a failure in the industrial machine or information indicating the “degree of failure”. The “failure information” may include information indicating that the industrial machine is in a normal state. The “degree of failure” means the seriousness of the failure. As the “degree of failure”, either the maximum value or the minimum value may be limited. The “degree of failure” may be a continuous amount or a discrete amount. The operator can determine whether to replace or repair the target component immediately or to perform the next maintenance work according to the “degree of failure”.

  In the following description, the failure prediction system 1 used for predicting a failure of the robot 2 will be described. However, those skilled in the art will recognize that the present invention is equally applicable to any other industrial machine.

  A robot 2 illustrated in FIG. 1 is a 6-axis vertical articulated robot in which each joint is driven by a motor. The robot 2 is connected to the robot control device 3 by a known communication means. The robot control device 3 creates a command for the robot 2 according to the control program.

  The robot control device 3 is a digital computer having a CPU, a ROM, a RAM, a nonvolatile memory, and an interface connected to an external device. As shown in FIG. 1, the robot control device 3 includes a failure determination unit 31.

  The failure determination unit 31 determines a failure of the robot 2 using a known failure diagnosis method. The failure determination unit 31 determines the presence or absence of the failure of the robot 2 or the degree of failure independently of the failure information created by the failure prediction system 1. For example, when the disturbance torque detected by the torque sensor or the amplitude of the vibration of the output data of the sensor exceeds a predetermined threshold, the failure determination unit 31 determines that a failure has occurred. Alternatively, the failure determination unit 31 may determine that a failure of the robot 2 has occurred based on internal data of control software stored in the robot control device 3. As described above, the failure determination unit 31 determines a failure based on various factors. The determination result by the failure determination unit 31 is input to a determination data acquisition unit 51 of the machine learning device 5 described later.

  In another embodiment, the machine learning device 5 is configured such that failure information is input to the determination data acquisition unit 51 in response to an operator's input operation that discovers or learns the failure of the robot 2. May be.

  The failure prediction system 1 further includes a sensor 11 that detects the state of the robot 2 or the surrounding environment. The sensor 11 is at least one of a force sensor, a torque sensor, a vibration sensor, a sound collection sensor, an image sensor, a distance sensor, a temperature sensor, a humidity sensor, a flow rate sensor, a light quantity sensor, a pH sensor, a pressure sensor, a viscosity sensor, and an odor sensor. One may be included. Data output from the sensor 11 (hereinafter, simply referred to as “output data”) is input to the state observation unit 52 of the machine learning device 5.

  The machine learning device 5 learns the failure condition of the robot 2. In one embodiment, the machine learning device 5 may be a digital computer that is connected to the robot 2 via a network and is separate from the robot control device 3.

  In another embodiment, the machine learning device 5 may be built in the robot control device 3. In this case, the machine learning device 5 performs machine learning using the processor of the robot control device 3. In another embodiment, the machine learning device 5 may exist on a cloud server.

  As shown in FIG. 1, the machine learning device 5 includes a determination data acquisition unit 51, a state observation unit 52, and a learning unit 53.

  The determination data acquisition unit 51 acquires determination data from the failure determination unit 31. The determination data is input from the determination data acquisition unit 51 to the learning unit 53 and is used when the machine learning device 5 learns the failure condition. The determination data is data that determines the presence or absence of a failure or the degree of failure. The determination data may not include data indicating that there is a failure, that is, that the robot 2 is in an abnormal state.

  The state observation unit 52 observes a state variable as an input value of machine learning while the robot 2 is operating or stationary. In an embodiment in which the machine learning device 5 is connected to the robot 2 and the sensor 11 via a network, the state observation unit 52 acquires a state variable via the network.

  The state variable may include output data of the sensor 11. The state variable may include internal data of control software that controls the robot 2. The internal data may include at least one of torque, position, speed, acceleration, jerk, current, voltage, and estimated disturbance value. The estimated disturbance value is a disturbance value estimated by an observer based on, for example, a torque command and speed feedback.

  The state variable may include calculation data obtained based on output data or internal data. The calculation data may be acquired using at least one of frequency analysis, time frequency analysis, and autocorrelation analysis. Of course, the calculation data may be obtained using simpler calculations, such as coefficient multiplication or differential integration.

  The learning unit 53 learns the failure condition according to the training data set created based on the combination of the state variable output from the state observation unit 52 and the determination data output from the determination data acquisition unit 51. The training data set is data in which state variables and determination data are associated with each other.

  An example of a learning process in the machine learning device 5 will be described with reference to FIG. When learning is started, in step S201, the state observation unit 52 acquires state variables including output data, internal data, calculation data, and the like. In step S <b> 202, the determination data acquisition unit 51 acquires determination data based on the determination result by the failure determination unit 31.

  In step S203, the learning unit 53 learns the failure condition according to the training data set created based on the combination of the state variable acquired in step S201 and the determination data acquired in step S202. The processes in steps S201 to S203 are repeatedly executed until the machine learning device 5 sufficiently learns the failure condition.

In one embodiment, the learning unit 53 of the machine learning device 5 may learn a failure condition according to a neural network model. FIG. 3 shows an example of a neural network model. Neural network comprises l neurons _{x 1, x 2, x 3} , ···, an input layer comprising a x _l, m neurons _{y 1, y 2, y 3} , ···, a y _m An intermediate layer (hidden layer) and an output layer including _n neurons z ₁ , z ₂ , z ₃ ,. In FIG. 3, although only one intermediate layer is shown, two or more intermediate layers may be provided. The machine learning device 5 (neural network) may use a general-purpose computer or processor. However, when GPGPU (General-Purpose computing on Graphics Processing Units) or a large-scale PC cluster is applied, the machine learning device 5 (neural network) performs higher-speed processing. It is possible.

  The neural network learns failure conditions associated with the failure of the robot 2. The neural network uses state variables observed by the state observation unit 52 and so-called supervised learning according to a training data set created based on a combination of the determination data acquired by the determination data acquisition unit 51. The relationship with failure occurrence, that is, failure conditions are learned. Supervised learning is a model that learns features in these datasets by giving a large number of sets of input and result (label) data to the learning device, and estimates the result from the input, that is, the relationship Can be obtained inductively.

  Alternatively, the neural network can accumulate only the state variables when there is no failure, that is, when the robot 2 is operating normally, and can learn the failure condition by so-called unsupervised learning. For example, when the failure frequency of the robot 2 is extremely low, an unsupervised learning method will be effective. Unsupervised learning is to provide only a large amount of input data to the learning device to learn how the input data is distributed, and compress the input data without providing corresponding teacher output data. -This is a method for learning a device that performs classification and shaping. It is possible to cluster the features in those datasets among similar people. Using this result, output can be predicted by assigning an output so as to optimize it by setting a certain standard. In addition, there is an intermediate problem setting between unsupervised learning and supervised learning called semi-supervised learning, in which only a part of the input and output data sets exist, and the others are input only. This is the case with data.

  FIG. 4 is a diagram for explaining an example of a learning period in an unsupervised learning technique. Here, the horizontal axis indicates time (elapsed time), and the vertical axis indicates the degree of failure. As shown in FIG. 4, the unsupervised learning method described above is based on a learning period of a certain period, for example, several weeks, starting from immediately after the robot 2 is shipped or immediately after maintenance. Only when the state variable is updated and defined as normal. After that, the state variable is not updated, and the abnormality detection is performed by outputting the “degree of failure” from the output result output from the neural network based on the distance from the normal model and performing only the abnormality determination. You can do it.

  In this embodiment, for example, it is also effective to use a neural network called a recurrent type in order to model time-series data with temporal correlation. A recurrent neural network (RNN) does not form a learning model using only the state at the current time, but also uses the internal state at the previous time. A recurrent neural network can be handled in the same way as a general neural network by developing a time axis network. Here, there are various types of recurrent neural networks. As an example, a simple recursive network (Elman Network) will be described.

  FIG. 5 is a diagram for explaining an example of a recurrent type neural network. FIG. 5 (a) shows a time axis expansion of the Elman network, and FIG. 5 (b) shows an error back propagation method (back propagation). : Backpropagation (Backpropagation) indicates the back propagation time through time (BPTT: Back Propagation Through Time). Here, backpropagation can be applied to the structure of the Elman network as shown in FIG.

  However, in the Elman network, unlike a normal neural network, as shown in FIG. 5 (b), an error propagates back in time, and this backpropagation is referred to as backpropagation through time (BPTT). Call. By applying such a neural network structure, it is possible to estimate an output model based on the transition of the input so far, for example, whether the estimated output value is a certain abnormal value or not. It can be used for the relationship with the occurrence.

  When performing failure prediction, which will be described later, in response to a state variable input to the input layer of the neural network, the output layer outputs information indicating the presence or absence of a failure corresponding to the failure information described above or “degree of failure”. Note that the possible value of the “degree of failure” may be a value in which either the maximum value or the minimum value is limited, a continuous amount, or a discrete amount.

  According to the machine learning device and the machine learning method according to the above-described embodiment, it is possible to learn a more accurate failure condition according to the actual use situation than the failure condition based on the determination data output from the determination data acquisition unit 51. Thereby, even when the factors leading to the failure are complicated and it is difficult to preset the failure condition, it is possible to predict the failure with high accuracy.

  In one embodiment, when the determination data acquisition unit 51 acquires determination data representing a failure of the robot 2, the learning unit 53 captures the determination data from the time when the failure occurred until the acquisition of each determination data. The failure condition may be updated by weighting according to the length. Here, it is estimated that the shorter the time from when the determination data is acquired until the failure actually occurs, the closer to the state where the failure occurs. Therefore, if the determination data is weighted according to the elapsed time since the training data set was acquired, the failure condition can be effectively learned.

  In one embodiment, the learning unit 53 may learn a failure condition according to a training data set created for a plurality of robots 2. The learning unit 53 may acquire a training data set from a plurality of robots 2 used at the same site, or training data collected from a plurality of robots 2 that operate independently at different sites. You may learn a failure condition using a set. In addition, the robot 2 that collects the training data set can be added to the target on the way, or conversely, can be removed from the target.

  Next, as examples of methods for sharing (sharing) training data sets of a plurality of robots 2, three examples will be given below, but it goes without saying that other methods can be applied. First, a first example is a method of sharing a neural network model so as to be the same. For example, for each weight coefficient of the network, a difference between the robots 2 is transmitted using a communication unit. It is reflected. As a second example, by sharing the input and output data sets of the neural network, the weight of the learning device 5 can be shared. Furthermore, as a third example, a state is prepared by preparing a certain database, accessing it, and loading a model of a more appropriate neural network (with a similar model).

  FIG. 6 is a block diagram illustrating an example of a failure prediction system according to another embodiment. The failure prediction system 1 includes a failure prediction device 4 that creates failure information of the robot 2 using a result learned by the machine learning device 5.

  The failure prediction device 4 includes a state observation unit 41 and a failure information output unit 42. The state observing unit 41 functions in the same manner as the state observing unit 52 described with reference to FIG. 1, and acquires a state variable reflecting the state of the robot 2 and the surrounding environment. The failure information output unit 42 responds to the input of the state variable via the state observation unit 41 based on the result learned by the learning unit 53 of the machine learning device 5 according to the training data set, and the failure information of the robot 2. Is output.

  As shown in FIG. 6, the robot control device 3 can include a notification unit (failure information notification unit) 32. The notification unit 32 notifies the operator of the failure information output by the failure information output unit 42. The manner in which the failure information is notified is not particularly limited as long as the operator can know it. For example, the presence or absence of a predicted failure or the degree of failure may be displayed on a display device (not shown), or a warning sound may be generated according to the content of failure information.

  7 and 8 are diagrams for explaining examples of index values (first to fourth examples) indicating the degree of failure in the failure prediction system according to the embodiment. Here, in FIG. 7A, FIG. 7B, FIG. 7C, and FIG. 8, the horizontal axis indicates time, and the vertical axis indicates the degree of failure. First, as shown in FIG. 7 (a), for example, in the first example, the index value indicating the “degree of failure” is set to increase as the failure approaches, and the index value obtained by learning is used as it is. The failure information output unit 42 can output the failure information. Further, as shown in FIG. 7B, for example, in the second example, a threshold value is provided for the above-described index value, and whether there is a failure such as abnormal if it is greater than or equal to the threshold value, normal if it is less than the threshold value. The failure information output unit 42 may output the information to be expressed as failure information. Further, as shown in FIG. 7 (c), for example, in the third example, a plurality of threshold values (threshold value 1 to threshold value 3) are provided for the above-described index value, and the levels (fault level 1 to fault value) are divided for each threshold value. The failure information output unit 42 may output level 4) as failure information.

  As shown in FIG. 8, for example, in the fourth example, a relationship between the above-described index value and time until failure is obtained based on data (teacher data) that led to a plurality of failures, and based on that, A first threshold value for satisfying that the time before the time determined in the first predetermined period is obtained from the time when the failure occurs is obtained. Further, a second threshold value is set for satisfying that the time after the time determined in the second predetermined period is traced back from the time when the failure occurs. Then, when satisfying at least one of the index value being less than the first threshold and the index value being greater than or equal to the second threshold, the index value itself, or a level obtained by dividing the index value by the threshold, The failure information output unit 42 can also output the failure information. In this case, the threshold value can be determined, for example, by setting the threshold value so that past teacher data satisfies all the conditions, or by providing a margin with a margin if necessary. In addition, a threshold value can be set so as to allow a determination error within a certain probability.

  Next, an example of failure prediction executed using the result learned by the machine learning device will be described with reference to FIG. In step S501, the state observation unit 41 acquires a current state variable including, for example, output data from the sensor 11. In step S502, the failure information output unit 42 outputs failure information corresponding to the state variable acquired in step S501 based on the learning result of the machine learning device 5 described above. When the failure prediction system 1 includes the notification unit 32, a step of notifying the operator of failure information may be executed after step S502.

  The failure prediction by the failure prediction device 4 described with reference to FIG. 9 may be performed when the robot 2 executes a predetermined specific operation. Alternatively, the processes in steps S501 to S502 may be continuously executed while the robot 2 is operating or stationary. Alternatively, failure prediction may be performed periodically at a predetermined time.

  In one embodiment, the machine learning by the machine learning device 5 may be executed in parallel with the execution of the failure prediction by the failure prediction device 4. In that case, at the same time that the failure prediction device 4 creates the failure information, the machine learning device 5 learns based on the determination data acquired through the operation of the failure determination unit 31 or the operator and the state variable at that time. The unit 53 relearns the failure condition.

  Although an embodiment of machine learning using a neural network has been described, machine learning may be performed according to other known methods such as genetic programming, functional logic programming, support vector machine, and the like. Again, in this specification, the term “industrial machine” means various machines including industrial robots, service robots, and machines controlled by computer numerical control (CNC) devices. As described above.

  Although various embodiments of the present invention have been described above, those skilled in the art will recognize that the functions and effects intended by the present invention can be realized by other embodiments. In particular, the components of the above-described embodiments can be deleted or replaced without departing from the scope of the present invention, or known means can be further added. It is obvious to those skilled in the art that the present invention can be implemented by arbitrarily combining features of a plurality of embodiments explicitly or implicitly disclosed in the present specification.

DESCRIPTION OF SYMBOLS 1 Failure prediction system 2 Robot 3 Robot control apparatus 4 Failure prediction apparatus 5 Machine learning apparatus 11 Sensor 31 Failure determination part 32 Notification part 41 State observation part 42 Failure information output part 51 Determination data acquisition part 52 State observation part 53 Learning part

Claims (13)

A machine learning device for learning conditions associated with a failure of an industrial machine,
It includes at least one of output data of a sensor that detects the state of the industrial machine or the surrounding environment, internal data of control software that controls the industrial machine, and calculation data obtained based on the output data or the internal data A state observation unit for observing state variables during operation or stationary of the industrial machine;
A determination data acquisition unit for acquiring determination data for determining the degree of failure of the industrial machine;
According training data set that is created based on a combination of the state variables and the decision data, Bei example and a learning unit that learns a condition associated with a failure of the industrial machinery,
The learning unit obtains the determination data included in the training data set from the occurrence of the failure until the determination data is acquired when the determination data acquisition unit acquires the determination data representing the failure of the industrial machine. Update the above condition by weighting according to the length of time that goes back.
A machine learning device characterized by that. The learning unit is configured to learn the condition according to the training data set created for a plurality of industrial machines.
The machine learning device according to claim 1. The learning unit learns a normal state only during a certain period, and then detects the occurrence of a failure by the determination data acquisition unit.
The machine learning apparatus according to claim 1, wherein the machine learning apparatus is a machine learning apparatus. A failure prediction device for predicting a failure of the industrial machine, comprising the machine learning device according to any one of claims 1 to 3 ,
Based on the result learned by the learning unit according to the training data set, in response to the current input of the state variable, a failure information output unit that outputs failure information indicating the presence or absence of the industrial machine or the degree of failure Further comprising
A failure prediction apparatus characterized by that. The learning unit re-learns the condition according to an additional training data set created based on the combination of the current state variable and the determination data.
The failure prediction apparatus according to claim 4 . The machine learning device is connected to the industrial machine via a network,
The state observation unit acquires the current state variable via the network;
The failure prediction apparatus according to claim 4 or 5 , characterized in that: The machine learning device exists on a cloud server,
The failure prediction apparatus according to claim 6 . The machine learning device is built in a control device that controls the industrial machine,
The failure prediction apparatus according to any one of claims 4 to 6 , wherein the failure prediction apparatus is characterized in that: The learning result by the machine learning device is shared by a plurality of the industrial machines.
The failure prediction apparatus according to any one of claims 4 to 8 , wherein the failure prediction apparatus is characterized in that: The failure prediction apparatus according to any one of claims 4 to 9 ,
A sensor for outputting the output data;
A failure information notification unit for notifying an operator of the failure information,
Failure prediction system characterized by that. The time when the failure information is notified to the operator by the failure information notification unit is earlier than the time determined in the first predetermined period retroactively from the time when the failure occurs,
The failure prediction system according to claim 10 . The time when the failure information is notified to the operator by the failure information notifying unit is earlier than the time determined in the first predetermined period, going back from the time when the failure occurs, and going back from the time when the failure occurs. , After a time determined by a second predetermined period longer than the first predetermined period,
The failure prediction system according to claim 11 . A machine learning method for learning conditions associated with an industrial machine failure,
It includes at least one of output data of a sensor that detects the state of the industrial machine or the surrounding environment, internal data of control software that controls the industrial machine, and calculation data obtained based on the output data or the internal data Observing state variables while the industrial machine is operating or stationary,
Obtaining determination data that determines the degree of failure of the industrial machine,
Learning a condition associated with a failure of the industrial machine according to a training data set created based on the combination of the state variable and the determination data ;
The condition associated with the failure of the industrial machine is learned because when the determination data representing the failure of the industrial machine is acquired, the determination data included in the training data set is changed from the occurrence of the failure to the determination data. Update the above condition by weighting according to the length of time going back to the time of acquisition,
A machine learning method characterized by that.

Images