JP2020149100A - Real-time controller and distributed control system and industrial machinery using the same - Google Patents

Abstract

To provide a real-time controller such that a control operation of the learning operation and the inference operation are the same, and a distributed control system and an industrial machinery using the same.SOLUTION: There is provided a real-time controller that is connected to a network that performs learning calculation of deep learning in a part thereof, takes in measurement data measured externally, and executes inference calculation of deep learning inference calculations. The real-time controller comprises: a setting holding unit for storing setting information including an input delay time of reference data obtained by learning calculation of deep learning, for the reference data among multiple measurement data; a time correction unit for selecting data other than the reference data at the latest timing after the input delay time has elapsed, for the data other than the reference data, in which the measurement data is stored, and the time when the input delay time elapses from the input time of the reference data is set as the input time of the reference data; and an arithmetic unit for executing inference operations of deep learning using the reference data and the selected data other than the reference data as a set of data.SELECTED DRAWING: Figure 3

Description

The present invention relates to a controller that executes real-time control, a real-time controller that uses real-time control using inference operations by deep learning, a distributed control system that uses the controller, and an industrial machine.

In recent years, deep learning technology has been used for the purpose of improving the accuracy of image recognition and predicting failures of industrial equipment. Especially in deep learning with supervised learning, an arbitrary arithmetic model can be obtained without mathematical design by executing learning arithmetic using input data and any corresponding teacher (correct answer) data as parameters. .. In addition, the calculation model can be updated by performing a re-learning calculation according to the input data and the error of the correct answer. Therefore, by using deep learning, it is possible to create a control calculation model (learning result) without considering the difference in the operating environment of the controlled object and complicated physical phenomena, and perform calculation based on this (hereinafter, inference calculation). It is possible, and it is expected that the system design man-hours will be reduced.

In general, machine control in an industrial machine such as an industrial device or a robot executes a control calculation based on information of an arbitrary sensor or actuator. This also applies to machine control using deep learning technology. For example, Japanese Patent Application Laid-Open No. 1 is a machine learning system provided with a learning unit, an inference unit, and a sensor, which is important for learning a machine learning system. It is characterized by the fact that the machine learning system performs high-precision inference operations by selecting and utilizing only various data.

JP-A-2010-55303

Mechanical control of industrial equipment and robots requires a high-speed control cycle from milliseconds to microseconds. On the other hand, in the learning operation and the inference operation, the data input / output path and the input / output method are often different depending on the calculation performance and the implementation form required for each operation, and the delay time at the time of input becomes large. Therefore, when the configuration of Patent Document 1 is applied to high-speed machine control, there is no consideration for the error between the data input delay time during the learning calculation and the inference calculation, which affects the high-speed control, and the learning calculation and the inference. The control operation at the time of calculation may not be the same.

From the above, it is an object of the present invention to provide a real-time controller having the same control operation during learning calculation and inference calculation, a distributed control system using the same, and an industrial machine.

From the above, in the present invention, there are a plurality of real-time controllers that are connected to a network that performs deep learning learning operations in a part of them, take in measurement data measured externally, and execute deep learning inference operations. Regarding the reference data among the measurement data of, the setting holding unit that stores the setting information including the input delay time of the reference data obtained by the learning calculation of deep learning, and the setting holding unit that stores the measurement data and the input delay time from the input time of the reference data. Is used as the input time of the reference data, and the time correction unit for selecting the data other than the reference data at the latest timing from the time when the input delay time has elapsed for the data other than the reference data, and the reference data are selected. A real-time controller characterized by having an arithmetic unit that executes inference operations for deep learning using data other than the reference data as a set of data. "

Further, in the present invention, the "real-time controller is a distributed control system configured by being connected to an upstream computer that executes deep learning via a network and a downstream computer that acquires the measurement data".

Further, in the present invention, it is defined as an "industrial machine configured by a distributed control system".

According to the present invention, in machine control using deep learning, the control operation during the learning operation and the inference operation can be made the same by correcting the data input / output delays during the learning operation and the inference operation under the same conditions. can do.

The figure which shows the structural example of the real-time controller which concerns on embodiment of this invention. The figure which shows the data structure which a real-time controller uses for communication. The figure which shows the detailed configuration example of the time correction part 16 provided in the real-time controller. A flowchart showing a setting procedure for correcting a delay time in a test run mode. The flowchart which shows the procedure of the delay time correction processing which the time correction unit 16 executes. The figure which shows the timing chart when the delay time correction processing is not performed with respect to the input data. The figure which shows the timing chart when the delay time correction processing is performed on the input data. The figure which shows the configuration example of the distributed control system 6 to which the real-time controller 1 is applied. The figure which shows the setting terminal screen which inputs the setting information necessary for the time correction part which a real-time controller has. The figure which shows the operation application example of the real-time controller 1 which concerns on Example 3 of this invention.

Hereinafter, examples of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a diagram showing a configuration example of a real-time controller according to an embodiment of the present invention.

The real-time controller 1 composed of computers communicates with the computers on the upstream side and the downstream side through a wired or wireless network. Further, for example, the upstream computer of the network executes the learning calculation of deep learning in advance such as during a trial run, and the real-time controller 1 executes an inference calculation reflecting the result of deep learning during normal operation.

The real-time controller 1 of FIG. 1 includes a communication control unit 12 that controls communication between the calculation unit 18, upstream and downstream computers, a time correction unit 16 that corrects the communication time, and upstream and upstream that control data transmission / reception. It includes at least downstream data control units 13 and 14.

More specifically, the communication control unit 12 includes an upstream communication port 10 for receiving or transmitting a transmission signal for upstream communication, and a downstream communication port 11 for receiving or transmitting a transmission signal for downstream communication. , Converts to transmission signal data of transmission / reception communication, or relays transmission signal. The data to be transmitted and received is controlled by the upstream and downstream data control units 13 and 14.

The time correction unit 16 obtains setting information necessary for processing of the time correction unit from the setting holding unit 15, determines reference data based on the preset setting, corrects the input delay time, and corrects the input delay time of the preset setting and the reference data. Based on the input delay time, data other than the reference data is selected to create a set of data consisting of the reference data and data other than the reference data.

The calculation unit 18 executes a calculation process using the data from the time correction unit 16, and at this time, the calculation measurement unit 17 measures the elapsed time of the start and end of the calculation time.

The upstream communication port 10 and the downstream communication port 11 are connected to the communication control unit 12, the communication control unit 12 is connected to the upstream data control unit 13 and the downstream data control unit 14, and the upstream data control unit 13 is the setting maintenance unit 15. The downstream data control unit 14 is connected to the time correction unit 16 and the calculation unit 18, the setting holding unit 15 is connected to the time correction unit 16, and the time correction unit 16 is connected to the calculation unit 18. The calculation unit 18 is connected to the communication control unit 12, the upstream data control unit 13, and the calculation measurement unit 17.

Here, the real-time controller 1 uses the structure shown in the packet 2 of FIG. 2 in order to send and receive data by communication. The packet 2 includes an address unit 20 indicating a destination when the real-time controller 1 communicates, and a data unit 21 for storing parameters used for calculation and setting information used by the time correction unit 16.

FIG. 3 is a configuration diagram showing details of the time correction unit 16 included in the real-time controller 1.

The time correction unit 16 accumulates the data interpretation unit 30 that interprets the data content of the packet received by the upstream data control unit 13 or the downstream data control unit 14, the time counting unit 31 that measures the time, and the received data. The input delay time of the reference data is corrected based on the settings stored in the data buffer unit 32, the reference selection unit 161 that selects the reference data based on the settings stored in the setting holding unit 15, and the settings stored in the setting holding unit 15. The delay correction unit 162 and the timing adjustment unit 163 that selects the input data and creates a data set based on the setting and the input delay time of the reference data stored in the setting holding unit 15 are provided, and the timing adjustment unit 163 provides the time. Pass the adjusted data to the calculation unit.

In this way, the data interpretation unit 30 is connected to the upstream data control unit 13, the downstream data control unit 14, and the data buffer unit, the time count unit 31 is connected to the data buffer unit 32, and the data buffer unit is the reference selection unit. The 161 and the timing adjustment unit 163 are connected, the reference selection unit 161 is connected to the delay correction unit 162, the delay correction unit 162 is connected to the timing adjustment unit 163, and the timing adjustment unit 163 is connected to the calculation unit 18. There is.

Here, the basic operation of the real-time controller 1 will be described.

First, the real-time controller 1 receives the settings required for the processing of the time correction unit 16 via the upstream communication port 10 in the upstream data control unit 13. The upstream data control unit 13 transfers the received setting information to the setting holding unit 15, and the setting holding unit 15 stores the transferred setting information. Here, the settings required for the processing of the time correction unit 16 are the reference data setting information for selecting the reference data to be the target of the delay compensation, and the delay time correction setting information for correcting the input delay time of the reference data. Information including.

At this time, both the upstream data control unit 10 and the downstream data control unit 11 can receive or transmit. Therefore, the real-time controller 1 constantly transmits and receives data having the structure of the packet 2 in the upstream direction and the downstream direction. In particular, the received data is accumulated in the data buffer unit 32 of the time correction unit 16, and the timing adjustment unit 163 selects only the data used in the calculation of the calculation unit 18. Further, the real-time controller 1 periodically executes a calculation in the calculation unit 18 based on the received data. At that time, the calculation unit 18 transfers the calculation result to the upstream data control unit 13 or the downstream data control unit 14. When the calculation result is transferred, the upstream data control unit 13 or the downstream data control unit 14 transfers the calculation result to the communication control unit 12 and converts it into a communication transmission signal.

Further, the real-time controller 1 includes a test run mode for setting the delay time correction process and a normal mode for executing the control calculation. In the test run mode, for example, a learning calculation for deep learning is executed by an upstream computer, and as a result of this test run, various setting information necessary for processing of the time correction unit 16 is obtained by the real-time controller 1 and the setting process is executed. In the normal mode, the real-time controller 1 executes an inference calculation using the learning model obtained as a result of the deep learning learning calculation, which is obtained in the test run mode and is required for the processing of the predetermined time correction unit 16. Delay time correction processing using various setting information is executed.

Next, the data delay time correction process executed by the time correction unit 16 included in the real-time controller 1 will be described. The setting procedure for correcting the delay time in the test run mode will be described with reference to FIG.

In FIG. 4, first, the real-time controller 1 is set to the test run mode in the processing step S40 in order to estimate the input delay and the calculation time of the data to be set. In the test run mode, the learning calculation of deep learning is executed by the upstream computer. Next, the operation information that is the result of the trial run in the processing step S41 is confirmed.

When the confirmation of the operation information is completed, next, in the processing step S42, the reference data as the reference for correcting the delay time is selected.

In the setting procedure (processing step S43 and processing step S44) of FIG. 4, two types of delay information are input for each reference data. The first is internal delay information, which is the time from when the data (reference data) to be input is acquired as a digital value to when it is transmitted to the real-time controller 1 and the data arrives. The second is external delay information, which is the time until the data to be input is converted from an analog value to a digital value.

Here, the internal delay information and the external delay information will be described in more detail. For example, assuming that the upstream computer is the control device for the entire plant and the downstream computer is the individual controller for each part of the plant, the process is performed by the downstream computer. The time from measuring the amount to converting the analog value to the digital value is the external delay information, and the time from the generation of the digital value to the recognition by the real-time controller 1 is the internal delay time in the downstream. Become.

These two types of information (reference data setting information for selecting reference data to be subject to delay compensation and delay time correction setting information for correcting the input delay time of reference data) are processed in processing step S43 and processing step S44. The setting is completed as soon as the input for each data is completed. Finally, in the processing step S45, the real-time controller 1 starts the operation in the normal mode.

FIG. 5 is a flowchart showing a procedure of the delay time correction process executed by the time correction unit 16 in the normal mode. Here, it is assumed that the data required for processing has already been stored in the setting holding unit 15. Although not shown in FIG. 5, in the processing of the real-time controller 1 in FIG. 5, an inference operation using a learning model obtained as a result of the learning operation of deep learning is executed.

In the processing step S50, the time correction unit 16 receives the data in the data interpretation unit 30. At this time, after confirming the data content, the data interpretation unit 30 reads the time at the time of data reception from the calculation / measurement unit 17 and transitions to the processing step S51. Next, in the processing step S51, the data buffer unit 32 accumulates the data received by the data interpretation unit 30 together with the data reception time, and transitions to the processing step S52.

In the processing step S52, the reference selection unit 161 sequentially confirms the data accumulated in the data buffer unit 32, and determines whether or not the data is a preset reference data. If it is determined in the determination of the processing step S52 that the data is not the reference data set in advance, the process returns to the processing step S50 and the reception of the data is continued.

If it is determined in the determination of the processing step S52 that the data is the reference data set in advance, the reference selection unit 161 notifies the delay correction unit 162 and transitions to the processing step S53. At this time, the reference data can be determined by designating the address unit 20 provided in the packet 2.

Next, in the processing step S53, the delay correction unit 162 starts the time measurement from the time when the reference data is received, and transitions to the processing step S54. In process step S54, it is determined whether the time being measured by the delay correction unit 162 has reached a preset time. At this time, if the time measured by the delay correction unit 162 in the determination of the processing step S54 does not reach the preset input delay time, the process stays at the processing step S54.

When the time measured by the delay correction unit 162 in the determination of the processing step S54 reaches the input delay time set in advance, the delay correction unit 162 notifies the timing adjustment unit 163 and transitions to the processing step S55. At this time, the preset input delay time is a total value of the internal delay information and the external delay information.

The time measurement from the processing step S53 to the processing step S54 is a procedure executed to correct the input delay time of the reference data set in advance. That is, the input delay time is corrected to an arbitrary input delay time by intentionally adjusting the time by the preset input delay time from the time when the reference data arrives.

In the processing step S55, the timing adjusting unit 163 selects data other than the reference data from the data stored in the data buffer unit 32 based on the input delay time measured in the processing step S54 and the preset setting.

Here, a mechanism for the timing adjustment unit 163 to select data will be described. As a premise, it is possible to set an arbitrary input delay time for data other than the reference data. When the timing adjusting unit 163 selects data from the data buffer unit 32, first, the data outside the range of the input delay time measured in the processing step S54 is ignored. At this time, as described above, since the data buffer unit 32 stores the received data and the data reception time together, it is possible to refer to the reception time of the data associated with this data. After that, the timing adjustment unit 163 calculates relative to the input delay time of the reference data by referring to the input delay time and the data reception time set for the data other than the reference data, and the data of the specific reception time. Select.

By the above processing, from the time-series data group of data other than the reference data, the data other than the reference data obtained at the latest time at the time determined by the input delay time set in advance for the reference data is selected. Become.

After that, the transition to the processing step S56 is performed, and the timing adjustment unit 163 creates a set of input data by combining the reference data and data other than the selected reference data, and transitions to the processing step S57.

In process step S57, the timing adjustment unit 163 determines whether or not all the preset data have been selected. If it is determined in the determination of the processing step S57 that all the preset data has not been selected, the process transitions to the processing step S56, and the timing adjustment unit 163 selects the data again.

If it is determined in the determination of the processing step S57 that all the preset data has been selected, the process transitions to the processing step S58. Finally, in the processing step S58, the timing adjustment unit 163 outputs the set of input data to the calculation unit 18.

FIG. 6 is a diagram showing a timing chart when the time correction unit 16 included in the real-time controller 1 does not perform the delay time correction processing on the input data.

Here, in the example shown in FIG. 6, data A and data B are used as input data, and the learning calculation of deep learning is executed by an arbitrary computer (in many cases, the upstream computer is used) in the test run mode, and the normal operation. The case where the inference calculation as the learning result is executed by the real-time controller 1 in the mode is shown in comparison. Here, it is attempted to explain that the result of the delay time differs between the test run and the normal operation.

In FIG. 6, regarding the data A, the inference performed by the real-time controller 1 is the overhead time T60 during the learning calculation performed by the upstream computer during the learning calculation (test run) and the inference calculation (actual operation). It shall be longer than the overhead T62 at the time of calculation. Further, regarding the data B, it is assumed that almost no overhead is applied during the learning operation and the inference operation. Note that T61 and T63 represent data A input time points during learning calculation and inference calculation, and T64 represents data B input time point.

Based on the above premise, data A is sampled in period T67, and data B is sampled in period T66. In this case, as soon as the data A is input, the latest sampling data B is selected and used in the calculation as a set of input data.

Specifically, at the time of the learning calculation, the data A is input to the real-time controller 1 at the time point T61 through the overhead T60 from the data measurement by the downstream computer. On the other hand, the data B is closest to the input time point T61 of the data A, and the data at the time point of T65, which is the latest sampling, is selected and used as a set of input data.

On the other hand, at the time of the inference calculation, the data A is input to the real-time controller 1 at the time point T63 via the overhead T62 from the data measurement by the downstream computer. On the other hand, for the data B, the data at the time of T64, which is the latest sampling closest to the input time point T63 of the data A, is selected and used as a set of input data.

That is, since the overhead of the data A is different between the learning operation and the inference operation, the timing at which the data B is selected is different, and the acquisition time of the data B in the set of input data between the learning operation and the inference operation is different. Will be different.

Next, a timing chart when the time correction unit 16 performs the delay time correction process on the input data will be described with reference to FIG. 7. Even in the case described here, the same premise as that described in FIG. 6 is assumed.

In FIG. 7, data A is treated as reference data and data B is treated as data other than reference data. Since the overhead of the reference data A is different between the learning operation and the inference operation, the delay time is adjusted in the inference operation with a short overhead and in the learning operation with a long overhead, so that the learning operation and the inference operation can be performed. , Achieve the reference data input time at the same timing. Further, data B other than the reference data is selected at the timing closest to the input time of the reference data A.

First, at the time of the learning calculation, the data A is input at the time point T61 through the overhead T60 from the data measurement by the downstream computer as in the case of FIG. At this time, the data B selects the data at the time point T611, which is the latest sampling closest to the time point T61, and uses it as a set of input data.

On the other hand, at the time of inference calculation by the real-time controller 1, in order to correct the delay time of the input data by the time correction unit 6, the data A is set as the reference data, and the delay time of the T610 is further set. Therefore, at the time of inference calculation and at the time of data arrival, the data is actually input at the timing T69 from the data measurement by the downstream computer through the overhead T68, but the input delay time is added only for the time T610. After that, it is assumed that the data is input at the time point T612, and at that time, the data at the time point T611, which is the latest sampling closest to the time point T612, is selected for the data B and used as a set of input data.

By performing such input delay correction, the overhead of data A is corrected to be the same at the time of learning calculation and at the time of inference calculation, and the timing at which data B is selected is also the same, so that at the time of learning calculation and inference calculation. It is possible to solve the problem that the acquisition timing of the data B is different among the sets of input data of.

According to the embodiment of the present embodiment, when the deep learning calculation operation and the inference operation are executed, the input data is input under the same conditions such as the input delay time even if the data input path and the calculation environment for execution are different. It is possible to create a set of.

FIG. 8 is a diagram showing a configuration example of a distributed control system 6 as an example of a system for actually operating the real-time controller 1 according to the second embodiment of the present invention.

The distributed control system 7 manages the entire distributed control system 7, a system management device 70, at least one real-time controller 1, at least one input / output relay device 71, at least one sensor 72, or at least one sensor 72. It is configured to include one actuator 73, or both at least one sensor 72 and at least one actuator 73, and a setting terminal device 74.

Here, the system management device 70 constitutes a computer system in which the upstream computer is used, and the input / output relay device 71 is used as a downstream computer. In this case as well, the system management device 70, which is an upstream computer of the network, executes the deep learning learning calculation during the trial run, and the real-time controller 1 executes the deep learning inference calculation during the actual operation.

In this way, the system management device 70 is connected to the real-time controller 1 or the input / output relay device 71, the real-time controller 1 is connected to another real-time controller 1 or the input / output relay device 71, and the input / output relay device 71 is connected to the input / output relay device 71. It is connected to the sensor 72 or the actuator 73, and the setting terminal device 74 is connected to the system management device 70.

Next, the details of each device will be described.

The system management device 70 includes at least one downstream communication port 11, a system management unit 701, a communication control unit 702, and a terminal connection port 703, and the system management unit 701 includes a communication control unit 702 and a terminal connection port. Connected to 703, the communication control unit 702 is connected to at least one downstream communication port 11.

The basic operation of the system management device 70 will be described. The system management device 70 manages the settings input by the setting terminal device 74 by the system management unit 701 and transfers them to the communication control unit 702, so that the time correction unit of the real-time controller 1 passes through the downstream communication port 11. The setting information required for 16 is transmitted. Here, the system management device 70 is assumed to be a server connected to a network, a general-purpose computer, a small computer board, or the like. Since the real-time controller 1 has been described in the first embodiment, the description thereof is omitted here.

The input / output relay device 71 includes at least one upstream communication port 10, at least one downstream communication port 11, communication control unit 710, input / output unit 711, and device connection port 712, and includes communication control unit 710. Is connected to at least one upstream communication port 10 and at least one downstream communication port 11, the input / output unit 711 is connected to the communication control unit 710 and the device connection port 712, and the device connection port 712 is the sensor 72 or the actuator 73. Is connected with.

The basic operation of the input / output relay device 71 will be described. The input / output relay device 71 is connected to the device connection port 712 and inputs / outputs to the sensor or actuator via the input / output unit 711. At this time, the input data is data obtained by transmitting the data obtained by the input / output unit 711 to the real-time controller 1 or the system management device 70 via the communication control unit 710 and the upstream communication port 10 or the downstream communication port 11. .. Further, the output data is data received from the real-time controller 1 or the system management device 70 via the upstream communication port 10 or the downstream communication port 11 and the communication control unit 710, and is output via the input / output unit 711. To do.

Next, the basic operation of the distributed control system 6 will be described.

In this embodiment, the real-time controller 1 executes an operation for obtaining a control command of the actuator 73 based on the input data of the plurality of sensors 72.

When using the distributed control system 6, first, in the setting terminal device 74, the real-time controller 1 inputs the setting information for executing the input delay correction.

FIG. 9 is a diagram showing a setting terminal screen for inputting necessary setting information to the time correction unit 16 included in the real-time controller 1. At this time, the setting terminal device 74 has a configuration as shown in FIG. The setting terminal device 74 includes a setting area 80 and a system log area 81. The setting area 80 specifies reference data and sets the delay time for each data in order to execute the input delay correction process. Further, in the system log area 81, the input setting information is transferred to the communication control unit 702 via the system management unit 701 and transmitted to the real-time controller 1.

As for the setting procedure at this time, as described in the first embodiment, after the test run mode is first executed, the operation information is confirmed in the system log area 81. Next, in the setting area, the internal delay information and the external delay information are input for each data to be corrected for the delay time. After that, the setting information input by the setting terminal device 74 is transferred to the real-time controller 1 via the system management device 70, and the real-time controller 1 further stores the received setting information in the setting holding unit 15.

Here, the input / output relay devices 71A and 71B to which the sensors 72A and 72B are connected continuously sample the data of the sensors 72A and 72B and transfer the data to the real-time controller 1. At this time, as described in FIG. 8, for convenience of explanation, the sensor 72A of the data A and the sensor 72B of the data B are used. Further, data A is used as reference data for input delay correction, and data B is used as data other than the reference data. When the real-time controller 1 receives the reference data A, the real-time controller 1 corrects the input delay time of the sensor 72A of the reference data A by the method described in the first embodiment. After that, among the data of the sensor 72B of the data B other than the plurality of reference data accumulated in the data buffer unit 30 provided in the real-time controller 1 based on the preset setting and the input delay time of the sensor 72A of the reference data A. Select one data.

At this time, the data B other than the reference data of the sensor 72B selected and the data of the sensor 72 of the reference data A are transferred to the calculation unit 18 as a set of input data. The calculation unit 18 performs a calculation based on the input data, and then transfers the calculation result to the input / output relay device 71C to which the actuator 73 is connected by communication. The input / output relay device 71C to which the actuator 73 is connected converts the received calculation result into a control signal to the actuator 73 by the input / output unit 711 and feeds it back.

According to the embodiment of the present embodiment, by applying the distributed control system 6 provided with the real-time controller 1 to an industrial machine such as an industrial device or a robot, the delay time of each data such as a sensor or an actuator input to a control calculation can be obtained. It is possible to specify the combination exactly.
Since the effects when the others are applied to the real-time controller 1 are the same as those in the first embodiment, the description thereof will be omitted.

FIG. 10 is an operational application example of the real-time controller 1 according to the third embodiment of the present invention.

The calculation unit 18 included in the real-time controller 1 includes a calculation execution unit 91, a calculation parameter storage unit 92, and a calculation management unit 93.

The calculation execution unit 91 is connected to the time correction unit 16, the upstream data control unit 10, and the downstream data control unit 11, and the calculation parameter storage unit 92 is connected to the calculation execution unit 91 and the communication control unit 12, and the calculation management unit 93. Is connected to the calculation execution unit 91 and the communication control unit 12.

At this time, the calculation execution unit 91 executes an arbitrary calculation. Further, the calculation parameter storage unit 92 stores the coefficients and parameters required for the calculation executed by the calculation execution unit 91, and changes and updates the contents via communication.

As shown in FIG. 10, the real-time controller 1 is operated by being connected to another real-time controller 1. At this time, when the real-time controller 1 is used alone, if the arithmetic performance for the target arithmetic processing is insufficient, a plurality of real-time controllers 1 are connected as shown in FIG. 10 and the arithmetic processing is distributed and implemented. So, we will reinforce the calculation performance. Specifically, the arithmetic management unit 93 included in the real-time controller 1 shares the progress status of the arithmetic processing and the arithmetic result with the arithmetic management unit 93 of the other real-time controller 1 via communication. That is, one type or a plurality of types of arithmetic processing having high required performance is executed in cooperation with the plurality of real-time controllers 1.

Further, the cloud server 90 transfers the arithmetic parameters related to the arithmetic processing to the arithmetic parameter storage unit 92 at an arbitrary timing via communication. Further, the real-time controller 1 continuously transmits the calculation result and the operation status to the cloud server 90, and the cloud server 90 reconsiders and updates the calculation parameters.

According to the third embodiment, the real-time controller 1 can update the calculation contents and expand the calculation performance according to the purpose of the calculation.

1: Real-time controller 10: Upstream communication port 11: Downstream communication port 12: Communication control unit 13: Upstream data control unit 14: Downstream data control unit 15: Setting holding unit 16: Time correction unit 161: Reference selection unit 162: Delay correction Unit 163: Timing adjustment unit 17: Arithmetic measurement unit 18: Arithmetic unit 2: Packet 20: Address unit 21: Data unit 30: Data interpretation unit 31: Time count unit 32: Data buffer unit 70: System management device 701: System management Unit 702: Communication control unit 703: Terminal connection port 71: Input / output relay device 710: Communication control unit 711: Input / output unit 712: Device connection port 72: Sensor 73: Actuator 74: Setting terminal device 80: Setting area 81: System Log area 90: Cloud server 91: Calculation execution unit 92: Calculation parameter storage unit 93: Calculation management unit

Claims (12)

A part of it is a real-time controller that is connected to a network that performs deep learning learning operations, takes in measurement data measured externally, and executes deep learning inference operations.
Regarding the reference data among the plurality of the measurement data, a setting holding unit that stores the setting information including the input delay time of the reference data obtained by the learning calculation of the deep learning, and the setting holding unit that stores the measurement data and inputs the reference data. The time when the input delay time has elapsed from the time is set as the input time of the reference data, and for the data other than the reference data, the data other than the reference data at the timing closest to the time when the input delay time has elapsed is selected. A real-time controller including a time correction unit and a calculation unit that executes an inference calculation of the deep learning by using the reference data and data other than the selected reference data as a set of data. The real-time controller according to claim 1.
A real-time controller characterized in that a learning calculation of deep learning is executed during a trial run of the network, and an inference calculation based on a learning model obtained as a result is executed by the calculation unit during actual operation. The real-time controller according to claim 1 or 2.
One or more upstream communication ports, one or more downstream communication ports, an upstream data control unit for transmitting / receiving data to the upstream communication port, and data transmission / reception to the downstream communication port. A downstream data control unit for connecting, a communication control unit connected to the upstream communication port, the downstream communication port, the upstream data control unit, and the downstream data control unit, and a connection between the upstream data control unit and the downstream data control unit. It is provided with the time correction unit, the calculation unit connected to the time correction unit, the communication control unit, the upstream data control unit, and the downstream data control unit.
For the data acquired by the upstream data control unit or the downstream data control unit, the time correction unit determines the reference data based on the preset setting, and the time correction unit determines the reference data based on the preset setting. The data input delay time is corrected, and the data other than the reference data is selected by the time correction unit based on the preset and the input delay time of the reference data, and the reference data and the reference data other than the reference data are selected. A real-time controller characterized by creating a set of data for. The real-time controller according to claim 3.
The setting holding unit is characterized in that the setting information required for the processing of the time correction unit is transferred by any means, or the previous period time correction unit included in the real-time controller processes based on the result of the calculation unit. Real-time controller to do. The real-time controller according to claim 4.
The setting information required for the processing of the time correction unit includes data designation information as a reference when sampling a plurality of input data required for the calculation of the calculation unit, and a plurality of inputs required for the calculation of the calculation unit. A real-time controller characterized by being data delay correction information. The real-time controller according to any one of claims 3 to 5.
The time correction unit is connected to the upstream data control unit and the downstream data control unit, and is connected to the data interpretation unit and the data interpretation unit that interpret the data acquired by the upstream data control unit or the downstream data control unit. , A data buffer unit for accumulating the acquired data, a time counting unit connected to the data buffer unit to measure time, and a reference selection unit connected to the data buffer unit to select reference data from the acquired data. Timing of connecting to the reference selection section and correcting the input delay time when the reference data is acquired, and the timing of connecting to the delay correction section and selecting specific data from the data buffer section. A real-time controller characterized by having an adjustment unit. The real-time controller according to claim 6.
The data interpretation unit interprets the received data content, the data buffer unit stores the data, and the reference selection unit notifies the delay correction unit when it receives the reference data selected based on the preset setting. The delay correction unit starts time measurement of the input delay time based on the notification from the reference selection unit, and the delay correction unit notifies the timing adjustment unit when a specific input delay time is reached based on a preset setting. The timing adjusting unit is a real-time controller characterized in that, based on a notification from the delay correction unit, the data is selected from the data buffer unit based on a preset setting, and a set of data is created. The real-time controller according to any one of claims 1 to 7 is connected to an upstream computer that executes the deep learning and a downstream computer that acquires the measurement data via a network. Distributed control system. The distributed control system according to claim 8.
A distributed control system characterized in that a plurality of the real-time controllers are connected on a network and one type or a plurality of types of operations are executed by the calculation unit included in the plurality of real-time controllers. The distributed control system according to claim 8 or 9.
The upstream computer is a system management device that manages the system, and the downstream computer is connected to at least one sensor or at least one actuator, or at least one sensor and at least one actuator. A distributed control system that is an input / output relay device that controls input / output to connected sensors or actuators, and is provided with a setting terminal device that inputs system logs and setting information on the network. The distributed control system according to claim 9.
The system management device receives the learning calculation result of deep learning via a server connected to the network and further transmits it to the real-time controller, and the real-time controller receives the learning result and information transferred from the input / output relay device. A distributed control system characterized by performing calculations based on. An industrial machine configured by the distributed control system according to any one of claims 8 to 10.

Images