Training device and inference device

Abstract

A training device (1A) includes: a data acquisition unit (12) that acquires time-series sensor data for tasks being carried out; a data storage unit (13) that stores data for each individual task acquired by the data acquisition unit (12); an identification time determination unit (14) that determines a time suitable for identifying a predetermined abnormality; and a training unit (15a) that trains an abnormality identification model by selecting data having a feature at the time determined by the identification time determination unit (14).

Description

Learning and inference devices

[0001] The present disclosure relates to a learning device and an inference device for diagnosing abnormalities when a robot performs a task.

[0002] Machine learning is being used to diagnose abnormalities during operation based on time-series data from control equipment such as industrial robots. In order to distinguish not only whether something is normal or abnormal but also the type of abnormality, a method has been proposed in which a large amount of normal-class data is acquired to learn a normal model, and data diagnosed as abnormal using the normal model is re-diagnosed using a different method (see, for example, Patent Document 1).

[0003] Japanese Patent Application Laid-Open No. 2018-24055

[0004] In the method of Patent Document 1, the acquired data is used for learning without selection, so in order to learn an accurate anomaly identification model based on robot data that varies from task to task, it is necessary to prepare a large amount of sample data for each type of anomaly being targeted.

[0005] The present disclosure has been made in view of the above, and aims to provide a learning device that can reduce the amount of data for each type of abnormality used in learning to identify abnormalities.

[0006] In order to solve the above-mentioned problems and achieve the objectives, the learning device according to the present disclosure has a data acquisition unit that acquires time-series sensor data during work, a data storage unit that saves data for each work trial acquired by the data acquisition unit, an identification time determination unit that determines a time suitable for identifying a predetermined abnormality, and a learning unit that selects data that has characteristics at the time determined by the identification time determination unit and learns an anomaly identification model.

[0007] The learning device according to the present disclosure has the effect of being able to reduce the amount of data for each type of anomaly used in learning to identify anomalies.

[0008] FIG. 1 is a block diagram showing the configuration of a learning device according to a first embodiment; FIG. 2 is a flowchart showing the procedure of an operation performed by a learning unit included in the learning device according to the first embodiment; FIG. 3 is a block diagram showing the configuration of an inference device according to the first embodiment; FIG. 4 is a block diagram showing the configuration of a learning device according to the second embodiment; FIG. 5 is a block diagram showing the configuration of a learning device according to the third embodiment;

[0009] A learning device and an inference device according to an embodiment will be described in detail below with reference to the drawings.

[0010] 1 is a block diagram showing the configuration of a learning device 1A according to embodiment 1. The learning device 1A includes a sensor 11, a data acquisition unit 12, a data storage unit 13, an identification time determination unit 14, and a learning unit 15a.

[0011] The sensor 11 is attached to the robot and measures data during work at a certain period. An example of the sensor 11 is a force F in the three orthogonal axial directions of the robot's hand. ｘ , F ｙ , F ｚ and the moment M around the three orthogonal axes ｘ , M ｙ , M ｚ It is a force sensor that measures the value of

[0012] The sensor 11 measures data when the robot is operated under normal conditions designed by the user and under conditions that intentionally cause an abnormality to be identified. Examples of abnormalities to be identified include abnormalities in parts such as partial loss or foreign matter contamination, or abnormalities in the robot's working position relative to surrounding equipment. In Figure 1, the data measured by the sensor 11 is shown as "experimental conditions for abnormality class."

[0013] The data acquisition unit 12 acquires time-series sensor data during work. For example, the data acquisition unit 12 acquires data measured by the sensors 11 at a certain cycle as time-series data. The cycle at which the data acquisition unit 12 acquires data does not need to be the same as the measurement cycle of the sensors 11. The data acquisition unit 12 may simultaneously acquire data measured by multiple types of sensors 11, such as data measured by a force sensor and data on the positions of each joint measured by an encoder of the robot. If the measurement cycles of the sensors 11 differ, the data acquisition unit 12 may perform resampling processing to acquire data with aligned timestamps.

[0014] The data storage unit 13 stores data for each task attempt acquired by the data acquisition unit 12. For example, the data storage unit 13 stores the time-series data for each task attempt acquired by the data acquisition unit 12 together with a correct answer label indicating whether the corresponding attempt is normal or abnormal. The data storage unit 13 is realized, for example, by a semiconductor memory.

[0015] The identification time determination unit 14 determines a predetermined time suitable for identifying an abnormality. For example, the identification time determination unit 14 determines the optimal identification time T ｏｐｔ (i, c) is calculated. ｏｐｔ A specific method for calculating (i, c) is to use the p-value of a t-test, which is known as a statistical testing method. In this method, the discrimination time determination unit 14 first selects data x of time k as a reference from a certain number of sample data for each class c. ｋ Probability distribution q(c|x ｋ Then, the discrimination time determination unit 14 calculates the probability distribution q(c'|x ｋ ) is the probability distribution q(c|x ｋ The probability that the probability of exceeding the threshold is obtained as the p-value of the t-test. In this case, the data interval used for evaluation may be taken as the period from the start of the work to a certain time k, and the identification time determination unit 14 may take data from time k-ΔT to time k using a sliding window with a time width ΔT. The optimal identification time Tｏｐｔ (i, c) can be the time when the p value is smallest in sample data i. Alternatively, a threshold ε is set, and the time when the p value becomes smaller than the threshold ε for the first time is the optimal discrimination time T ｏｐｔ (i, c). In this way, the identification time determination unit 14 may determine a predetermined time suitable for identifying an anomaly based on a probability distribution obtained when data is observed from the start of work until a certain time. The identification time determination unit 14 may also determine a predetermined time suitable for identifying an anomaly based on a probability distribution obtained when data is observed within a sliding window of a certain time width.

[0016] The learning unit 15a learns an anomaly identification model by selecting data having characteristics at the time determined by the identification time determination unit 14. The characteristics here refer to feature quantities for identifying anomalies, and appear in time-series data at times when the difference in values ​​between classes is large and the difference in values ​​within the same class is relatively small. The learning unit 15a may select data to add to learning based on the uncertainty of the time determined by the identification time determination unit 14.

[0017] 2 is a flowchart showing the procedure of the operation performed by the learning unit 15a of the learning device 1A according to embodiment 1. In step S1, the learning unit 15a determines the number n of learning samples for each class c to be classified. ｃ For example, the learning unit 15a initializes (class 0: number of normal samples, class 1: number of samples of abnormality type 1, class 2: number of samples of abnormality type 2, ...) = (n ０ , n １ , n ２ , ...).

[0018] In step S2, the learning unit 15a calculates the optimal discrimination time T for each learning sample data i of class c by the above-mentioned method. ｏｐｔ (i, c) is obtained from the identification time determination unit 14 .

[0019] In step S3, the learning unit 15a calculates all the optimal discrimination times T ｏｐｔ The average value of (i, c) T ａｖｅ (c) Calculate the average value T ａｖｅ (c) is expressed by the following formula (1).

[0020]

[0021] In step S4, the learning unit 15a calculates the average value T ａｖｅ The uncertainty evaluation value U(c) of class c in (c) is calculated. The uncertainty here means that due to variations in data within the same class, it is not possible to reliably determine that the class is c from the value at that time. A specific method for calculating the uncertainty evaluation value U(c) is to use the probability distribution q(c|x ｋ ) or the average p-value. Selection criteria such as expected error reduction, which have been proposed in active learning research, can also be used.

[0022] In step S5, the learning sample of the class with the largest uncertainty evaluation value U(c) is stored in the data storage unit 13, and the learning unit 15a selects and adds the learning sample of the class with the largest uncertainty evaluation value U(c) from the sample data that has not been used for learning, and n ｃ , n ｃ ←n ｃ +1. As more samples are added, the uncertainty estimate U(c) calculated using the above method becomes smaller and smaller.

[0023] In step S6, the learning unit 15a performs machine learning using the updated number of training samples and evaluates the resulting anomaly identification model. In FIG. 2 , the operation of step S6 is indicated by the phrase "evaluate the anomaly identification model using the current number of training samples." The learning unit 15a can use any machine learning method for classification problems for learning. For example, the learning unit 15a can use logistic regression to learn a model in which the posterior probability of the correct class when certain time-series data is obtained is 1 and the posterior probability of other classes is 0. Alternatively, the learning unit 15a can use a neural network such as a recurrent neural network (RNN) to learn a model in which the output layer values ​​are (1, 0, 0, ...) for class 0 data, (0, 1, 0, ...) for class 1 data, and (0, 0, 1, ...) for class 0 data. To evaluate the anomaly identification model, the learning unit 15a can use the accuracy rate of anomaly identification for sample data that has not been used for learning, or the cross entropy error, which is known as a loss function for classification problems.

[0024] In step S7, the learning unit 15a determines whether the performance of the anomaly identification model is sufficient. For this determination, a conditional expression can be used, such as that the accuracy rate is equal to or greater than a predetermined value, or that the value of a loss function such as a cross-entropy error is less than a predetermined value. If the learning unit 15a determines that the performance is sufficient (Yes in S7), it ends the learning and saves the anomaly identification model. If the learning unit 15a determines that the performance is not sufficient (No in S7), it returns to step S2 and considers the next sample to be added to learning.

[0025] 3 is a block diagram showing the configuration of an inference device 2A according to embodiment 1. The inference device 2A infers an anomaly from time-series data during work using the anomaly identification model learned by the learning device 1A. The inference device 2A includes a sensor 11, a data acquisition unit 12, a learning unit 15a, and an anomaly identification unit 16a.

[0026] The sensor 11, data acquisition unit 12, and learning unit 15a in Figure 3 are the same as those included in Figure 1. The anomaly identification unit 16a calls the anomaly identification model obtained by the learning unit 15a for time-series data acquired in the same format as during learning, and identifies the type of anomaly, including whether it is normal or abnormal. For example, when the anomaly identification unit 16a uses the anomaly identification model learned using the above-mentioned logistic regression, it can determine the class with the highest posterior probability value as the identification result. The identification result may be output at a certain period while the robot is working, or may be output at a specific time point during work or when the work is completed.

[0027] As described above, the learning device 1A according to the first embodiment can reduce the amount of data for each type of anomaly used in learning to identify anomalies. The inference device 2A according to the first embodiment can infer the type of anomaly using the anomaly identification model obtained by learning. Furthermore, the inference device 2A can identify the type of anomaly during work.

[0028] 4 is a block diagram showing the configuration of a learning device 1B according to embodiment 2. The learning device 1B includes a sensor 11, a data acquisition unit 12, a data storage unit 13, an identification time determination unit 14, and a learning unit 15b. The sensor 11, the data acquisition unit 12, the data storage unit 13, and the identification time determination unit 14 are the same as those described in embodiment 1, and therefore will not be described in embodiment 2.

[0029] The learning process flow performed by the learning unit 15b is the same as the process flow performed by the learning unit 15a of the learning device 1A according to the first embodiment. That is, the learning unit 15b executes the learning operation according to the procedure of the flowchart in FIG. 2. However, in the second embodiment, when adding a learning sample in step S5 of FIG. 2, the learning unit 15b does not add a sample to be added to learning from the sample data already stored in the data storage unit 13 at the start of learning, but instead operates the robot anew to acquire data for one task trial. Therefore, the learning unit 15b outputs a class, i.e., an experimental condition, for acquiring new data to be added to learning, based on the time uncertainty determined by the identification time determination unit 14.

[0030] Regarding the method of determining the class for which additional data is to be acquired, the learning unit 15b may determine the class for which the uncertainty evaluation value U(c) obtained in step S4 of Fig. 2 is the largest as the class for which additional data is to be acquired. Alternatively, in consideration of the risk that sample data obtained by newly operating the robot may be an outlier, the learning unit 15b may regard this problem as a multi-armed bandit problem in which "if you select and draw one arm from a slot machine, you will receive a reward based on a certain probability," and apply the ε-greedy algorithm or the UCB (Upper Confidence Bound) algorithm, which are known as selection algorithms for multi-armed bandit problems.

[0031] The configuration and processing contents of the inference device that performs inference using the anomaly identification model obtained by the learning device 1B of embodiment 2 are the same as those of the inference device 2A of embodiment 1, and therefore will not be described again.

[0032] As described above, according to the second embodiment, learning is started when only a small amount of sample data is obtained, and the number of experiments for acquiring new learning data can be reduced.

[0033] 5 is a block diagram showing the configuration of a learning device 1C according to embodiment 3. The learning device 1C includes a sensor 11, a data acquisition unit 12, a data storage unit 13, and a learning unit 15c. The sensor 11, the data acquisition unit 12, and the data storage unit 13 are the same as those described in embodiment 1, and therefore will not be described in embodiment 3.

[0034] The learning device 1C according to the third embodiment does not include the discrimination time determination unit 14. The learning unit 15c outputs a condition for acquiring new data to be added to learning based on the uncertainty of a certain time of the task. In the learning process flow performed by the learning unit 15c, the learning unit 15c calculates the optimal discrimination time T ｏｐｔ The average value T of the time based on (i, c) ａｖｅ Instead of calculating (c), we use the average value T ａｖｅAlternatively, the learning unit 15c sets the completion time of one task trial as the average value T ａｖｅ The subsequent learning process flow is the same as that in the second embodiment, and therefore a description thereof will be omitted.

[0035] The configuration and processing contents of the inference device that performs inference using the anomaly identification model obtained by the learning device 1C of embodiment 3 are the same as those of the inference device 2A of embodiment 1, and therefore will not be described again.

[0036] As described above, according to the third embodiment, even when training an anomaly identification model at a pre-designed identification time, training can be started from a state where only a small amount of sample data has been obtained, thereby reducing the number of experiments for acquiring new training data.

[0037] 6 is a block diagram showing the configuration of a learning device 1D according to embodiment 4. The learning device 1D has a sensor 11, a data acquisition unit 12, a data storage unit 13, an identification time determination unit 14, and a learning unit 15d. The sensor 11, the data acquisition unit 12, the data storage unit 13, and the identification time determination unit 14 are the same as those described in embodiment 1, and therefore will not be described in embodiment 4.

[0038] The fourth embodiment differs from the first embodiment in that time series data is acquired at an operating position that is offset from a reference position as an anomaly, and a regression model of the offset amount is learned. In the fourth embodiment, the anomaly is an offset of the operating position from the reference position, and the anomaly identification model learned by the learning unit 15d is a regression model of the offset amount. The learning process flow performed by the learning unit 15d is the same as the process flow in FIG. 2. The learning device 1D classifies the direction in which the offset is applied, for example, the translation direction or rotation direction with respect to the surface on which the assembly work is performed, as class c, and stores the time series data for each trial of the work together with the correct value of the offset amount in the data storage unit 13. The identification time determination unit 14 determines the optimal identification time T ｏｐｔ Calculate (i, c).

[0039] 2, the learning unit 15d performs machine learning to obtain a model that identifies the offset amount and the direction of the offset amount using a machine learning method for regression problems. For example, the learning unit 15d uses Gaussian Process Regression (GPR), which is a method used to estimate a nonlinear function y=f(x), to obtain a model that identifies the offset amount and the direction of the offset amount at the identification time T ａｖｅ Alternatively, the learning unit 15d can use a neural network such as an RNN to learn a model in which the first node of the output layer represents the translation offset amount in the X direction, the second node represents the translation offset amount in the Y direction, and so on.

[0040] In the fourth embodiment, the learning device 1D can also output experimental conditions for acquiring additional data from the learning unit 15d in combination with the method described in the second embodiment. For example, the learning device 1D can determine which offset direction, i.e., which sample data of class c, to add using the method described in the second embodiment, and can determine how much offset to apply in that direction to perform an experiment using Bayesian optimization, which is known as an algorithm that updates an acquisition function based on the results of Gaussian process regression and selects the next search point.

[0041] 7 is a block diagram showing the configuration of an inference device 2B according to embodiment 4. The inference device 2B performs inference using the anomaly identification model obtained by the learning device 1D. More specifically, the inference device 2B uses the regression model learned by the learning device 1D to infer an offset amount from time-series data during operation. The inference device 2B includes a sensor 11, a data acquisition unit 12, a learning unit 15d, and an anomaly identification unit 16d. The anomaly identification unit 16d calls the regression model obtained by the learning unit 15d for time-series data acquired in the same format as during learning, and outputs an offset direction identification result and an estimated value of the offset amount.

[0042] As described above, according to the fourth embodiment, even in the case of identifying an offset of an operating position that may occur during repetitive work as an abnormality, it is possible to reduce the amount of data used for learning and further to infer the current offset amount during work. In this embodiment, as in the third embodiment, it is also possible to configure without the identification time determination unit 14.

[0043] 8 is a block diagram showing the configuration of an inference device 2C according to embodiment 5. Like the inference device 2B according to embodiment 4, the inference device 2C includes a sensor 11, a data acquisition unit 12, a learning unit 15d, and an abnormality identification unit 16d, and further includes a motion correction unit 17 that corrects the current motion position based on the inferred offset amount.

[0044] The fifth embodiment is characterized in that correction to normal operation is performed during work, i.e., before the robot completely stops due to an error, based on an abnormality identification result using the anomaly identification model obtained by the learning device 1D according to the fourth embodiment. Since the output of the anomaly identification unit 16d is an identification result of the offset direction relative to the operation position where the work is normally performed and an estimated value of the offset amount, it is possible to correct the operation by adding a correction amount that cancels out the offset to the robot's operation command position. To prevent the error state from worsening due to excessive correction, the operation correction unit 17 may be designed to output a relaxed correction amount corresponding to the magnitude of uncertainty in the estimated value, which is represented by the magnitude of the kernel function in Gaussian process regression, for example.

[0045] As described above, according to embodiment 5, the current motion position can be corrected to a normal position during work. Furthermore, even when using an anomaly identification model obtained by the learning devices 1A, 1B, and 1C according to embodiments 1 to 3 that learn only the abnormality class, the inference device 2C can use the motion corrector 17 to select an action such as discarding an abnormal part or returning the hand position to the origin in response to the anomaly identification result.

[0046] 9 is a diagram showing a processor 91 in a case where at least some of the functions of the sensor 11, data acquisition unit 12, identification time determination unit 14, and learning unit 15a included in the learning device 1A according to embodiment 1 are realized by the processor 91. In other words, at least some of the functions of the sensor 11, data acquisition unit 12, identification time determination unit 14, and learning unit 15a may be realized by the processor 91 executing a program stored in memory 92. The processor 91 is a CPU (Central Processing Unit), a processing system, an arithmetic system, a microprocessor, or a DSP (Digital Signal Processor). FIG. 9 also shows the memory 92.

[0047] When at least some of the functions of the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a are realized by the processor 91, the functions are realized by the processor 91, software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 92. The processor 91 reads and executes the program stored in the memory 92, thereby realizing at least some of the functions of the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a.

[0048] When at least some of the functions of the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a are realized by the processor 91, the learning device 1A includes the processor 91 and a memory 92 for storing a program that results in the execution of at least some of the steps executed by the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a. It can also be said that the program stored in the memory 92 causes a computer to execute at least some of the procedures or methods executed by the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a.

[0049] The memory 92 may be, for example, a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), an EEPROM (registered trademark) (Electrically Erasable Programmable Read-Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disk).

[0050] 10 is a diagram showing a processing circuit 93 in a case where at least some of the functions of the sensor 11, the data acquisition unit 12, the discrimination time determination unit 14, and the learning unit 15a included in the learning device 1A according to embodiment 1 are realized by the processing circuit 93. In other words, at least some of the functions of the sensor 11, the data acquisition unit 12, the discrimination time determination unit 14, and the learning unit 15a may be realized by the processing circuit 93.

[0051] The processing circuitry 93 is dedicated hardware, and may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0052] Some of the functions of the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a may be realized by dedicated hardware that is separate from the remaining units.

[0053] With regard to the multiple functions of the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a, some of the multiple functions may be realized by software or firmware, and the remaining multiple functions may be realized by dedicated hardware. In this way, the multiple functions of the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15a can be realized by hardware, software, firmware, or a combination of these.

[0054] At least some of the functions of the sensor 11, the data acquisition unit 12, the learning unit 15a, and the anomaly identification unit 16a of the inference device 2A according to the first embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 91. The processing circuit is a processing circuit equivalent to the processing circuit 93.

[0055] At least some of the functions of the sensor 11, data acquisition unit 12, identification time determination unit 14, and learning unit 15b of the learning device 1B according to the second embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 91. The processing circuit is a processing circuit equivalent to the processing circuit 93.

[0056] At least some of the functions of the sensor 11, the data acquisition unit 12, and the learning unit 15c of the learning device 1C according to the third embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 91. The processing circuit is a processing circuit equivalent to the processing circuit 93.

[0057] At least some of the functions of the sensor 11, the data acquisition unit 12, the identification time determination unit 14, and the learning unit 15d of the learning device 1D according to embodiment 4 may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 91. The processing circuit is a processing circuit equivalent to the processing circuit 93.

[0058] At least some of the functions of the sensor 11, the data acquisition unit 12, the learning unit 15d, and the anomaly identification unit 16d of the inference device 2B according to the fourth embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 91. The processing circuit is a processing circuit equivalent to the processing circuit 93.

[0059] At least some of the functions of the sensor 11, the data acquisition unit 12, the learning unit 15d, the anomaly identification unit 16d, and the operation correction unit 17 of the inference device 2C according to the fifth embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 91. The processing circuit is a processing circuit equivalent to the processing circuit 93.

[0060] The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or different embodiments may be combined with each other, or part of the configuration may be omitted or modified within the scope of the gist of the invention.

[0061] 1A, 1B, 1C, 1D Learning device, 2A, 2B, 2C Inference device, 11 Sensor, 12 Data acquisition unit, 13 Data storage unit, 14 Identification time determination unit, 15a, 15b, 15c, 15d Learning unit, 16a, 16d Anomaly identification unit, 17 Operation correction unit, 91 Processor, 92 Memory, 93 Processing circuit.

Claims

1. A learning unit that learns an anomaly identification model based on time-series sensor data that has characteristics at times suitable for identifying anomalies. A learning device comprising:

2. A discrimination time determination unit that determines a time suitable for discrimination based on a probability distribution when data is observed up to a certain time. The learning device according to claim 1 , further comprising:

3. A discrimination time determination unit that determines a time suitable for discrimination based on a probability distribution when data is observed in a sliding window of a certain time width. The learning device according to claim 1 , further comprising:

4. The learning unit selects data to be added to learning based on the uncertainty of the time suitable for the discrimination.

2. The learning device according to claim 1 .

5. The learning unit outputs a condition for newly acquiring data to be added to learning based on the uncertainty of the time suitable for the identification.

2. The learning device according to claim 1 .

6. 6. An inference device that infers an anomaly from the time-series sensor data using an anomaly identification model learned by the learning device according to claim 1.

7. the anomaly is an offset of the operating position from a reference position; The anomaly identification model learned by the learning unit is a regression model of an offset amount.

6. The learning device according to claim 1, wherein the learning device is a learning device for learning a plurality of data.

8. 8. An inference device that infers an offset amount from the time-series sensor data using the regression model learned by the learning device according to claim 7.

9. a motion correction unit that corrects the current motion position based on the inferred offset amount; 9. The inference device according to claim 8, further comprising: