JP6809882B2 - Machine learning devices for learning fan failure prediction, devices including machine learning devices, and machine learning methods - Google Patents

Description

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

Conventionally, for example, a CNC (Computer Numerical Control) device, a robot controller, a servo amplifier, or the like is equipped with a cooling fan (fan motor). From the viewpoint of preventive maintenance, these fans are required to detect signs and promote maintenance before a failure occurs.

By the way, as failure (defective) modes due to fan life, for example, sticking and bearing life are known, but both failure modes include startup failure and steady rotation failure. Here, the start failure is a mode in which the normally rotating fan does not rotate at the next power-on (ON), and the steady rotation failure is a mode in which the normally rotating fan suddenly stops rotating. It is a mode called.

In this way, if the fan becomes defective, for example, the cooling capacity of the fan may decrease, and the temperature of the CNC device, robot controller, servo amplifier, etc. provided with the fan may increase, resulting in deterioration of performance or failure. It may occur.

By the way, conventionally, as a device for detecting and predicting a failure of a cooling fan such as a computer, the current for driving the fan is monitored, the rotation speed is detected from the periodic change in the current with rotation, and the detected value is predetermined. It has been proposed that a failure is determined when the value becomes smaller than the value of, and the rotation speed data is taken into the processor in time series to support the failure analysis prediction (see, for example, Patent Document 1). It is described that as a means for monitoring the state of the fan, there are a method of monitoring the rotation speed itself with a tachometer, a method of monitoring the temperature with a thermistor, and the like.

Further, conventionally, as a failure prediction and warning system for a fan of a cooling device, the temperature of a fan motor is monitored and an alarm is issued when a predetermined temperature abnormality occurs, so that a failure can be predicted before a complete failure. Those have been proposed (see, for example, Patent Document 2). Here, it is also described that there are two modes of fan failure, one is a sudden one in which lint or the like is entangled and the other is due to the life of the motor.

Further, conventionally, as a failure prediction system of a device (for example, an image forming device) composed of a large number of components, machine learning of the situation at the time of failure for many parameters related to the failure is performed to extract a plurality of failure modes. , A system has been proposed in which failure prediction is performed by comparing with the values of each parameter in the current operating state (see, for example, Patent Document 3).

Further, conventionally, as a system for predicting the occurrence of an abnormal situation in advance with a machine tool, a system has been proposed in which data in a normal machining operation is used for learning and a failure is predicted from the change over time (for example, patent documents). See 4).

Japanese Unexamined Patent Publication No. 09-184859 Japanese Patent No. 3011486 JP 2013-109483 JP-A-2015-203646

As described above, for example, the CNC device robot controller and the servo amplifier are provided with a fan (fan motor) for cooling, and for example, there are various factors in the failure mode of the fan.

Attempts have been made to acquire information related to such fan failure modes and detect fan failures using a predetermined algorithm, but the reality is that it is difficult to predict fan failures with high accuracy. Is.

According to an example of the first embodiment according to the present invention, it is a machine learning device that learns fan failure prediction, and is a state observation unit that observes state variables related to the operation of the fan, and a failure of the fan. Learning to create a learning model for predicting a fan failure based on a label acquisition unit that acquires a label relating to the presence or absence, the state variable observed by the state observation unit, and the label acquired by the label acquisition unit. A machine learning device equipped with a unit is provided.

According to another example of the first embodiment according to the present invention, it is a machine learner that learns fan failure prediction, and has a state observation unit that observes state variables related to the operation of the fan and the state observation unit. A machine learning device including a learning unit that generates a learning model of distribution and regularity based on the state variables observed by the unit is also provided.

According to the second embodiment of the present invention, the device includes the machine learning device of one example or another example of the first embodiment, and the device is provided with a device cooled by the fan.

According to an example of the third embodiment according to the present invention, it is a machine learning method for learning fan failure prediction, observing state variables related to the fan operation, and labeling the presence or absence of the fan failure. A machine learning method is provided that creates a learning model that predicts a failure of the fan based on the acquired and observed state variables and the acquired label.

According to another example of the third embodiment according to the present invention, it is a machine learning method for learning fan failure prediction, in which state variables related to the operation of the fan are observed, and the observed state variables are used. Based on this, a machine learning method for generating a learning model of distribution and regularity is provided.

According to the machine learning device, the device including the machine learning device, and the machine learning method according to the present invention, there is an effect that the failure of the fan can be predicted with high accuracy.

FIG. 1 is a diagram schematically showing a first embodiment of the machine learning device according to the present invention. FIG. 2 is a block diagram showing an example of a CNC device to which the machine learning device of the first embodiment is applied. FIG. 3 is a diagram schematically showing a second embodiment of the machine learning device according to the present invention. FIG. 4 is a diagram for explaining an example of the machine learning device shown in FIG. FIG. 5 is a diagram for explaining another example of the machine learning device shown in FIG.

Hereinafter, embodiments of a machine learning device for learning fan failure prediction according to the present invention, a device including the machine learning device, and a machine learning method will be described in detail with reference to the accompanying drawings. FIG. 1 is a diagram schematically showing a first embodiment of the machine learning device according to the present invention. For example, the fan does not rotate when the power is turned on and started next time. It shows an example of a machine learning device. As shown in FIG. 1, the machine learning device 2 of the first embodiment includes a state observation unit 21, a learning unit 22, an output utilization unit 23, and a label acquisition unit 24.

The state observation unit 21 is given input data given from the environment 1, for example, observed as state variables (state quantities) of the current, rotation speed, and temperature of the fan (fan motor) 100 that cools the CNC device 10. .. Specifically, the state observing unit 21 is, for example, the current and rotation speed of the fan 100 at the time of starting when the power is turned on, and the steady state in which a predetermined time has elapsed after the power is turned on and stable operation is performed. The current and the rotation speed of the fan 100 in the above are observed as state variables. Further, the state observing unit 21 observes, for example, the winding temperature of the fan 100, the ambient temperature of the fan 100, the air volume of the fan 100, the vibration of the fan 100, and the sound of the fan 100 as state variables.

Here, the current of the fan 100 can be detected as, for example, the current flowing through the fan 100 detected by the CNC device 10, and the rotation speed of the fan 100 can be detected, for example, by the output of a rotary encoder provided in the fan 100. it can. The winding temperature of the fan 100 corresponds to the temperature of the fan (motor) 100 itself, and can be detected by, for example, a temperature sensor attached to the fan 100. The ambient temperature, air volume, vibration, and sound of the fan 100 can be detected by corresponding sensors (that is, a temperature sensor, an air volume sensor, a vibration sensor, a sound sensor (microphone), and the like).

The label acquisition unit 24 acquires labels (teacher data), and acquires labels relating to the presence or absence of failure of the fan 100. The learning unit 22 performs "supervised learning", and predicts a failure of the fan 100 based on the state variables observed by the state observation unit 21 and the label (presence or absence of failure) acquired by the label acquisition unit 24. Create a learning model to do. The output utilization unit 23 receives the output of the learning unit 22, and displays, for example, that the fan 100 is likely to fail when it is determined that the fan 100 is likely to fail (for example). , Displayed on the operation display of the CNC device, or performed an alarm (for example, turning on a lamp) or prompting the replacement of the fan 100.

In the above, as the environment 1, the CNC device 10 installed in the machine tool in the factory has been described as an example, but this is not limited to the CNC device, and for example, a robot controller, a servo amplifier, and further. Needless to say, it may be various devices having a cooling function by the fan 100. Further, the number of fans 100 may be plural, and when the types are different, it is preferable to perform machine learning for the fans (motors) of each model number.

FIG. 2 is a block diagram showing an example of a CNC device to which the machine learning device of the first embodiment is applied, and the learning unit 22 is composed of an error calculation unit 221 and an error model update unit 222. In FIG. 2, the decision-making unit 25 is additionally drawn. As shown in FIG. 2, the CNC device 10 is provided with a fan 100 so as to blow air to cool the CNC device 10. As described above, the CNC device 10 may be a robot controller or a servo amplifier. Here, the sensor 200 is, for example, a temperature sensor that detects the ambient temperature of the fan 100, a microphone that detects the sound from the fan 100, or the like, but various other sensors (air volume sensor, vibration sensor, etc.) described above. ) Can also be provided in the vicinity of the fan 100.

As shown in FIG. 2, the learning unit 22 receives the output (state variable) and model (teacher data) of the state observation unit 21 and calculates the error, and the error calculation unit 221 and the output and error calculation of the state observation unit. It includes an error model update unit 222 that receives the output of the unit 221 and updates the error model related to the failure prediction of the fan 100. The label input to the error calculation unit 221 corresponds to the output of the label acquisition unit 24 of FIG. The decision-making unit 25 corresponds to, for example, the output utilization unit 23 in FIG. 1 described above, and defines information related to failure prediction of the fan 100 based on the output of the learning unit 22 (error model update unit 222).

The machine learning unit 2 can be applied with a general-purpose computer or processor, but for example, when a GPGPU (General-Purpose computing on Graphics Processing Units), a large-scale PC cluster, or the like is applied, processing can be performed at a higher speed. Becomes possible. Further, the machine learning device 2 can be connected to at least one other machine learning device, and a learning model created by the learning unit 22 of the machine learning device 2 can be connected to the machine learning device 2 with at least one other machine learning device. Can be exchanged or shared with each other.

As described above, according to the machine learning device of the first embodiment, for example, it is possible to predict with high accuracy a fan failure due to a sticking failure in which the fan does not rotate when the power is turned on and started next time.

FIG. 3 is a diagram schematically showing a second embodiment of the machine learning device according to the present invention. For example, a failure prediction of a fan due to a steady rotation failure that suddenly stops rotating when the fan is rotated and used. This is an example of a machine learning device that learns. As shown in FIG. 3, the machine learning device 2'of the second embodiment includes a state observation unit 21, a learning unit 22', and an output utilization unit 23'.

Similar to the first embodiment described above, the state observation unit 21 observes, for example, the current, the number of rotations, and the temperature of the fan 100 that cools the CNC device 10 as state variables as input data given from the environment 1. Given. That is, the state observation unit 21 is, for example, the current and rotation speed of the fan 100 at the time of starting when the power is turned on, and the fan in the steady state in which a predetermined time elapses after the power is turned on and stable operation is performed. Observe 100 currents and rotation speeds as state variables. Further, the state observing unit 21 observes, for example, the winding temperature of the fan 100, the ambient temperature of the fan 100, the air volume of the fan 100, the vibration of the fan 100, and the sound of the fan 100 as state variables. The sensor for observing each state variable is the same as that of the first embodiment, and the description thereof will be omitted.

The learning unit 22'generates a learning model of distribution and regularity based on the state variables observed by the state observation unit 21. Here, the learning unit 22'outputs the normal time score when the fan 100 is operating normally, and outputs the use time score when the fan 100 is actually used. The normal time score is obtained, for example, when the fan 100 is operated at a time when the possibility of failure is extremely low (a time when normal operation is compensated).

The output utilization unit 23'predicts the failure of the fan 100 in use based on the normal score and the use score output from the learning unit 22'. For example, the normal score is stored in the memory in advance. Then, by comparing the stored normal score with the use score output from the learning unit 22', a fan failure due to a steady rotation failure is predicted. That is, a learning model (normal score) is obtained by unsupervised learning based on the state variables (fan current, rotation speed, temperature, air volume, vibration, sound, etc.) when the fan 100 is rotating normally. It is generated and stored in memory, and its normal score is compared with the learning model (score during use) when it is actually used to predict fan failure due to steady-state rotation failure. The machine learning device 2'of the second embodiment can be connected to at least one other machine learning device as in the case of the first embodiment, and can be connected to at least one other machine learning device. The learning models created by the learning unit 22'of the machine learning device 2'can be exchanged or shared with each other.

Next, an example of the learning and determination method will be described. That is, as an example, the user designs a neural network model and performs learning using the data of the fan 100 at the normal time. Suppose that the output score of the machine learning device 2'has three elements, A, B, and C.

After the learning is completed, a defective sample (a broken fan) is prepared and input to the machine learning device 2'. The output score is compared with the normal score to determine the normal score range. Suppose that the normal score is determined according to the following condition 1.
A1 <A <A2, B1 <B <B2, C1 <C <C2 ... (Condition 1)

Here, when the score does not fall within the range of condition 1, for example, the score judgment device for determining the score sends a signal indicating an abnormality to the CPU, and the CPU receives a signal indicating an abnormality from the score determination device. Is displayed on the screen (display) as a warning to the upper management system, for example. Here, the machine learning device 2'is built in, for example, the CNA device 10, and the warning can be displayed on the operation display of the CNC device 10.

Here, unsupervised learning will be described. In the application example of unsupervised learning, there are many examples such as "input all data without preference and let the learner classify the input data", but in this second embodiment, only "data at normal time" is input. And learn. That is, in the second embodiment, in order to determine two ways of "presence or absence of fan failure" based on the input data, if the characteristics are learned using only one of the data, The other can be discriminated naturally. In addition, when all the data is input and classified without preference, it is not possible to learn a sufficient judgment device unless the data at the time of abnormality is added as the input data, but a sufficient amount for learning the judgment. Collecting anomalous data is difficult and impractical.

Therefore, in the second embodiment, unsupervised learning is applied in which only "normal data" is used as an input to learn the feature amount, and a normal score in the normal state is generated. The machine learning applied in the second embodiment is unsupervised learning because the "normal / abnormal" label is not applied in the learning process.

Hereinafter, the process of obtaining the output (score) from the input data will be described. FIG. 4 is a diagram for explaining an example of the machine learning device shown in FIG. 3, and conceptually shows the process of obtaining a score from input data. As shown in FIG. 4, in the second embodiment, it is assumed that the "neural network" is used as the machine learning device 2', but a general model of the neural network can be applied. it can.

As shown in FIG. 4, in the second embodiment, for example, "normal data Xn" by the fan 100, which is considered to be performing stable operation without aging deterioration, is used as the input X. It is given to the input part of the neural network. Here, the neural network includes an input unit, an intermediate layer, and an output unit, and the intermediate layer is composed of a plurality of layers. A feature amount (score) Z is output from the output unit of the neural network, and when "normal data" is given to the input unit, a set of normal scores can be obtained.

FIG. 5 is a diagram for explaining another example of the learner circuit shown in FIG. 3, and conceptually shows a self-encoder (autoencoder). As shown in FIG. 5, the neural network has a configuration in which an input unit, an intermediate layer, and an output unit, and an inverted input unit and the intermediate layer are added.

In the self-encoder shown in FIG. 4, input data (Xn) is given as input X to the input section of the neural network, and data output from the output section is taken out as output Y from the added intermediate layer and input section. To. As a result, the error between the input data Xn (input X) and the output data Yn (output Y) can be obtained as || X2n−Yn || ² . That is, the input feature amount can be extracted by the output unit by calculating the error by using the input data as it is as the correct label and learning so that the error is minimized. Then, by determining the error due to the input data during normal use and during actual use, it becomes possible to predict the failure of the fan due to the steady rotation failure with high accuracy.

When applying the neural network to the machine learning device 2'of the second embodiment, a general-purpose computer or processor can be used, but for example, when a GPGPU or a large-scale PC cluster is applied, the processing speed is higher. Will be possible. Further, the method of unsupervised learning is not limited to the above-mentioned one, and various methods such as non-hierarchical clustering by the k-means method or dimension compression in hierarchical clustering are applied. It goes without saying that you can do it.

Although the embodiments have been described above, all the examples and conditions described here are described for the purpose of assisting the understanding of the concept of the invention applied to the invention and the technology, and the examples and conditions described in particular are described. It is not intended to limit the scope of the invention. Nor does such description in the specification indicate the advantages and disadvantages of the invention. Although embodiments of the invention have been described in detail, it should be understood that various modifications, replacements and modifications can be made without departing from the spirit and scope of the invention.

1 Environment 2, 2'Machine learner 21 State observation unit 22, 22' Learning unit 23 Output utilization unit 24 Label acquisition unit 25 Decision-making unit 100 Fan motor 200 Sensor 221 Error calculation unit 222 Error model update unit

Claims (7)

A machine learner that learns fan failure prediction
A state observing unit that observes state variables related to the operation of the fan,
A learning unit that generates a learning model of distribution and regularity based on the state variables observed by the state observation unit is provided.
The learning unit
A normal score is output when the fan is operating normally, and a use score is output when the fan is in use.
Further, an output utilization unit for predicting a failure of the fan in use based on the normal score and the use score is provided.
The machine learning device
Learning the failure prediction of the fan due to steady rotation failure that stops rotating when the fan is rotated and used.
A machine learning device that features that. The state observation unit
Observe the current, rotation speed and temperature of the fan,
The machine learning device according to claim 1, wherein the machine learning device is characterized in that. The state observation unit
The current and rotation speed of the fan at startup, the current and rotation speed of the fan at steady time, the winding temperature of the fan, the ambient temperature of the fan, the air volume of the fan, the vibration of the fan, and the vibration of the fan. Observe the sound,
The machine learning device according to claim 2 , wherein the machine learning device is characterized in that. The machine learning device
Have a neural network,
The machine learning device according to any one of claims 1 to 3 , wherein the machine learning device is characterized in that. A device including the machine learning device according to any one of claims 1 to 4 .
The device is cooled by the fan.
A device characterized by that. The device
CNC device, robot controller, or servo amplifier,
The apparatus according to claim 5 . It is a machine learning method for learning fan failure prediction.
Observe the state variables related to the operation of the fan and
Based on the observed state variables, a learning model of distribution and regularity was generated.
Based on the normal score output when the fan is operating normally and the use score output when the fan is in use, the failure of the fan in use is predicted. ,
Learning the failure prediction of the fan due to steady rotation failure that stops rotating when the fan is rotated and used.
A machine learning method characterized by that.

Images