Determination system, learning system, determination method, learning method, and program

Abstract

The present disclosure addresses the problem of providing a determination system that generally determines an abnormal state on the basis of one-dimensional waveform data. A determination system (1) comprises an acquisition unit (11), a conversion unit (12), and a determination unit (14). The acquisition unit (11) acquires discontinuous one-dimensional waveform data (31) in which a start point and an end point are present. The conversion unit (12) converts the one-dimensional waveform data (31) into frequency data (33). The determination unit (14) determines, using a trained model (4) that is a neural network, whether or not the frequency data (33) indicates an abnormal state from a combination of the frequency data (33) and reference waveform data. The trained model (4) is generated by machine learning from training data. The training data includes frequency data indicating an abnormal state, and frequency data indicating a normal state.

Description

Determination system, learning system, determination method, learning method, and program

[0001] The present disclosure relates to a determination system, a learning system, a determination method, a learning method, and a program, and more particularly to a determination system, a learning system, a determination method, a learning method, and a program that use one-dimensional waveform data.

[0002] Patent Document 1 discloses an abnormality determination device that determines whether the state of a machine is normal or abnormal. The abnormality determination device in Patent Document 1 divides a detection signal indicating vibration from a vibration sensor into vibration waveforms for each component based on the detection signal from a rotation angle detection sensor. The abnormality determination device then compares the peak value, average intensity, and spectrum analysis results of the vibration waveform with thresholds or normal patterns and abnormal patterns to determine whether the state of the machine is normal or abnormal.

[0003] However, the anomaly determination device of Patent Document 1 requires that a threshold value for determining whether the state of a mechanical device is normal or abnormal, or normal patterns and abnormal patterns, be acquired in advance. That is, in the anomaly determination device of Patent Document 1, the criteria for determining whether the state of a mechanical device is normal or abnormal are unique to each mechanical device. Therefore, in the anomaly determination device of Patent Document 1, if the structure of the mechanical devices or the arrangement of the vibration sensors differ among multiple mechanical devices, using a uniform criteria for the vibration waveform may result in a decrease in determination accuracy.

[0004] Japanese Patent Application Publication No. 10-281859

[0005] The present disclosure aims to provide a determination system, a learning system, a determination method, a learning method, and a program that generally determine an abnormal state based on one-dimensional waveform data.

[0006] A determination system according to one aspect of the present disclosure includes an acquisition unit, a conversion unit, and a determination unit. The acquisition unit acquires discontinuous one-dimensional waveform data having a start point and an end point. The conversion unit converts the one-dimensional waveform data into frequency data. The determination unit uses a trained model to determine whether the frequency data indicates an abnormal state based on a combination of the frequency data and reference waveform data. The trained model is a neural network generated by machine learning from training data. The training data includes frequency data indicating an abnormal state and frequency data indicating a normal state.

[0007] A learning system according to one aspect of the present disclosure includes an acquisition unit, a conversion unit, a state acquisition unit, and a learning unit. The acquisition unit acquires a plurality of discontinuous one-dimensional waveform data each having a start point and an end point. The conversion unit converts the plurality of one-dimensional waveform data into a plurality of frequency data. The state acquisition unit acquires a plurality of state data indicating whether each of the plurality of frequency data indicates an abnormal state. The learning unit generates a trained model, which is a neural network, using training data. The training data includes the plurality of frequency data. The trained model receives as input first frequency data, which is one of the plurality of frequency data, and second frequency data, which is one of the plurality of frequency data and does not indicate an abnormal state, and outputs whether the first frequency data indicates an abnormal state.

[0008] A determination method according to one aspect of the present disclosure is executed by one or more processors. The determination method includes an acquisition step, a conversion step, and a determination step. In the acquisition step, discontinuous one-dimensional waveform data having a start point and an end point is acquired. In the conversion step, the one-dimensional waveform data is converted into frequency data. In the determination step, a trained model is used to determine whether the frequency data indicates an abnormal state based on a combination of the frequency data and reference waveform data. The trained model is a neural network generated by machine learning from training data. The training data includes frequency data indicating an abnormal state and frequency data indicating a normal state.

[0009] A program according to one aspect of the present disclosure causes one or more processors to execute the determination method.

[0010] A learning method according to one aspect of the present disclosure is executed by one or more processors. The learning method includes an acquisition step, a conversion step, a state acquisition step, and a learning step. In the acquisition step, a plurality of discontinuous one-dimensional waveform data each having a start point and an end point is acquired. In the conversion step, the plurality of one-dimensional waveform data is converted into a plurality of frequency data. In the state acquisition step, a plurality of state data indicating whether each of the plurality of frequency data indicates an abnormal state is acquired. In the learning step, training data is used to generate a trained model, which is a neural network. The training data includes the plurality of frequency data. The trained model receives as input first frequency data, which is one of the plurality of frequency data, and second frequency data, which is one of the plurality of frequency data and indicates a normal state, and outputs whether the first frequency data indicates an abnormal state.

[0011] A program according to another aspect of the present disclosure causes one or more processors to execute the learning method.

[0012] FIG. 1 is a functional block diagram of a determination system according to an embodiment. FIG. 2 is a schematic diagram of production equipment whose operation is to be determined by the determination system. FIG. 3 is an example of one-dimensional waveform data in the determination system. FIG. 4 is an example of frequency data corresponding to a normal state of production equipment in the determination system. FIG. 5 is an example of frequency data corresponding to an abnormal state of production equipment in the determination system. FIG. 6 is a schematic diagram showing learned data in the determination system. FIG. 7 is a functional block diagram of a learning system according to an embodiment. FIG. 8 is a schematic diagram showing a learned model in the learning system.

[0013] Hereinafter, a determination system, a learning system, a determination method, a learning method, and a program according to embodiments will be described in detail with reference to the drawings. However, each diagram described in the following embodiments is a schematic diagram, and the ratios of the sizes and thicknesses of the components do not necessarily reflect the actual dimensional ratios. Note that the configurations described in the following embodiments are merely examples of the present disclosure. The present disclosure is not limited to the following embodiments, and various modifications are possible depending on the design, etc., as long as the effects of the present disclosure can be achieved.

[0014] (Embodiments) (1) Overview A determination system 1 (see FIG. 1) is a system that determines whether the operation of a production facility 2 (see FIG. 2) is abnormal, for example, based on one-dimensional waveform data 31 (see FIG. 3) that indicates one of a plurality of operations repeatedly performed in the production facility 2. The production facility 2 is, for example, a molding machine that uses a mold 21. Specifically, the operation of the production facility 2 being abnormal refers to a slug defect in the mold 21. The slug defect refers to a situation in which, during a hole-punching process using the mold 21, a portion 24 of a punched material 23 (hereinafter referred to as "slug") is not properly discharged and remains in the mold 21. The slug defect refers to a situation in which the material 23 is not properly processed due to the presence of the slug 24 in the mold 21.

[0015] A sensor 22 is installed in the production facility 2. The sensor 22 outputs sensor data 32 indicating multiple operations repeatedly performed in the production facility 2. Specifically, the sensor 22 is a strain sensor that detects distortion of the mold 21. The sensor data 32 is data indicating a time series of distortion values ​​of the mold 21. The sensor data 32 includes one or more one-dimensional waveform data 31. Specifically, one piece of one-dimensional waveform data is a portion of the sensor data 32 indicating one of the operations repeatedly performed in the production facility 2. A portion from the start time to the end time of one press operation among multiple press operations corresponds to one piece of one piece of one-dimensional waveform data 31. The one-dimensional waveform data 31 is discontinuous data having a start point and an end point. Here, "discontinuous data" refers to data of a finite length. Furthermore, the one-dimensional waveform data 31 is data indicating a time series change of one value. In the embodiment, one value is the distortion value of the mold 21.

[0016] (2) Configuration of Determination System As shown in FIG. 1, the determination system 1 includes an acquisition unit 11, a conversion unit 12, a storage unit 13, and a determination unit 14.

[0017] The acquisition unit 11 acquires discontinuous one-dimensional waveform data 31 having a start point and an end point. Specifically, the acquisition unit 11 acquires sensor data 32 from the sensor 22 and extracts one or more pieces of one-dimensional waveform data 31 from the sensor data 32.

[0018] As shown in Fig. 3, the one-dimensional waveform data 31 is data that indicates a time-series change in one piece of data. Here, the one piece of data is the distortion value of the mold 21. Here, as shown in Fig. 3, one-dimensional data 31a when no slug lift-up occurs has a different waveform from one-dimensional data 31b when slug lift-up occurs.

[0019] The acquisition unit 11 acquires, for example, information indicating the start time and end time of each of a plurality of operations that are repeatedly performed from the production equipment 2, and extracts one or more pieces of one-dimensional waveform data 31. For example, if the production equipment 2 is a molding machine, the acquisition unit 11 acquires the press height or the rotation angle of the crank that operates the press, and extracts one or more pieces of one-dimensional waveform data 31 from the sensor data 32. Note that the sensor 22 may be linked to the operation of the production equipment 2, and output one piece of one-dimensional waveform data 31 each time the production equipment 2 performs each of a plurality of operations.

[0020] The converter 12 converts the one-dimensional waveform data 31 acquired by the acquirer 11 into frequency data 33. The frequency data 33 is data indicating the frequency distribution of the one-dimensional waveform data 31. More specifically, the frequency data 33 indicates the frequency spectrum of the one-dimensional waveform data 31 at each time from the start point to the end point.

[0021] In the first embodiment, the frequency data 33 is two-dimensional image data. Specifically, as shown in FIGS. 4 and 5 , the frequency data 33 is rectangular image data in an XY orthogonal coordinate system, and the brightness of each pixel indicates the frequency spectrum of the one-dimensional waveform data 31. For example, in the frequency data 33, the X coordinate indicates a time in the one-dimensional waveform data 31, the Y coordinate indicates a frequency, and the brightness of a pixel indicates the intensity of a frequency component corresponding to the Y coordinate at a time in the one-dimensional waveform data 31 corresponding to the X coordinate. Note that FIG. 4 shows frequency data 33a corresponding to one-dimensional data 31a. Also, FIG. 5 shows frequency data 33b corresponding to one-dimensional data 31b.

[0022] The converter 12, for example, performs a Fourier transform on a portion of the one-dimensional waveform data 31 that spans a predetermined time interval starting from the first time point, thereby creating a frequency distribution corresponding to the first time point. The predetermined time interval is, for example, 10 seconds, but is not limited to this and may be determined appropriately depending on the sampling rate and frequency distribution of the one-dimensional waveform data 31. The converter 12 then maps the intensity of the frequency component corresponding to the Y coordinate of each pixel having an X coordinate corresponding to the first time point in the frequency data 33 as the pixel brightness. The converter 12 performs the above process, treating each time point from the start point to the end point of the one-dimensional waveform data 31 as the first time point, thereby generating frequency data 33. The converter 12 stores the generated frequency data 33 in the storage unit 13.

[0023] The storage unit 13 is a storage medium that stores a plurality of frequency data 33 .

[0024] The determination unit 14 uses the trained model 4 to determine whether the frequency data 33 indicates an abnormal state from a combination of the frequency data 33 and the reference waveform data 33A (see FIG. 6 ). Here, "the frequency data 33 indicates an abnormal state" means that one operation corresponding to the frequency data 33 is in an abnormal state among multiple operations of the production equipment 2. In other words, "the frequency data 33 indicates an abnormal state" means that a sludge buildup occurred in one operation of the production equipment 2 corresponding to the frequency data 33.

[0025] The determination unit 14 generates or selects reference waveform data 33A based on the plurality of frequency data 33 stored in the storage unit 13. The reference waveform data 33A has characteristics similar to one or more frequency data 33 that indicate a normal state among the plurality of frequency data 33. The reference waveform data 33A is, for example, a representative value of two or more frequency data 33 among the plurality of frequency data 33 stored in the storage unit 13. For example, the reference waveform data 33A is average data of the three most recent frequency data 33 among the plurality of frequency data 33.

[0026] The determination unit 14 inputs the frequency data 33 and the reference waveform data 33A to the trained model 4 as two-dimensional image data. The trained model 4 receives the frequency data 33, which are two-dimensional images, and the reference waveform data 33A as input, and outputs information indicating whether the frequency data 33 is in an abnormal state. The trained model 4 is a neural network, and more specifically, a program for causing the determination unit 14 to realize a neural network. The trained model 4 is generated by machine learning using training data. A method for generating the trained model 4 will be described later.

[0027] The determination system 1 includes a computer system. The computer system is primarily composed of a processor and memory as hardware. The processor executes a program stored in the memory of the computer system to realize the functions of the determination system 1 of the present disclosure. The program may be pre-recorded in the memory of the computer system, provided via a telecommunications line, or provided in a non-transitory recording medium such as a memory card, optical disk, or hard disk drive that is readable by the computer system. The processor of the computer system is composed of one or more electronic circuits including a semiconductor integrated circuit (IC) or a large-scale integrated circuit (LSI). The integrated circuits, such as ICs and LSIs, are referred to by different names depending on the degree of integration, and include integrated circuits called system LSIs, very large-scale integrations (VLSIs), or ultra-large-scale integrations (ULSIs). Furthermore, field-programmable gate arrays (FPGAs), which are programmed after the LSI is manufactured, or logic devices that allow the reconfiguration of internal connections or internal circuit partitions of the LSI, can also be used as processors. The electronic circuits may be integrated into one chip or distributed across multiple chips. The chips may be integrated into one device or distributed across multiple devices. The computer system referred to here includes a microcontroller having one or more processors and one or more memories. Therefore, the microcontroller is also composed of one or more electronic circuits including a semiconductor integrated circuit or a large-scale integrated circuit.

[0028] Furthermore, it is not essential for the determination system 1 that multiple functions in the determination system 1 are concentrated in one housing, and the components of the determination system 1 may be distributed across multiple housings. Furthermore, at least some of the functions of the determination system 1, for example, some of the functions of the determination unit 14, may be realized by the cloud (cloud computing) or the like.

[0029] Conversely, at least some of the functions of the determination system 1, which are distributed among multiple components in the embodiment, may be integrated into a single housing.

[0030] (3) Trained Model The trained model 4 is, for example, a convolutional neural network in which the fully connected layer is excluded. As shown in FIG. 6 , the trained model 4 includes multiple convolutional layers 41 and multiple pooling layers 42.

[0031] Each of the multiple convolution layers 41 performs spatial filtering on the image by shifting a spatial filter whose size is smaller than the size of the image.

[0032] Each of the multiple pooling layers 42 reduces the size of the image by combining multiple pixels contained in a given region within the image into one.

[0033] The trained model 4 includes a first independent block 43a, a second independent block 43b, a combination layer 44, and an integrated block 43c.

[0034] The first independent block 43a includes one or more (one in FIG. 6 ) convolutional layers 41 and one or more (one in FIG. 6 ) pooling layers 42. The frequency data 33 is input to the first independent block 43a. The first independent block 43a outputs image data to the combination layer 44. That is, the first independent block 43a processes the frequency data 33 using one or more convolutional layers 41 and one or more pooling layers 42, and outputs the processing results.

[0035] The second independent block 43b includes one or more (one in FIG. 6 ) convolutional layers 41 and one or more (one in FIG. 6 ) pooling layers 42. The reference waveform data 33A is input to the second independent block 43b. The second independent block 43b outputs image data to the combination layer 44. That is, the second independent block 43b processes the reference waveform data 33A using one or more convolutional layers 41 and one or more pooling layers 42, and outputs the processing results.

[0036] The combination layer 44 combines the output of the first independent block 43a and the output of the second independent block 43b. For example, the combination layer 44 combines the image data output by the first independent block 43a and the image data output by the second independent block 43b. The combination layer 44 may create and output a difference image between the image data output by the first independent block 43a and the image data output by the second independent block 43b. The combination layer 44 may also combine the difference image with either or both of the image data output by the first independent block 43a and the image data output by the second independent block 43b.

[0037] The integration block 43c includes one or more (one in FIG. 6 ) convolution layers 41 and one or more (one in FIG. 6 ) pooling layers 42. The image data output by the combination layer 44 is input to the integration block 43c. The integration block 43c outputs data indicating whether or not the frequency data 33 indicates an abnormal state. The integration block 43c outputs, for example, image data having a spatial dimension of 1×1. The pixel value of one pixel included in the image data indicates whether or not the frequency data 33 indicates an abnormal state. In other words, the multiple convolution layers 41 of the trained model 4 are trained on processing results having a spatial dimension of 1×1.

[0038] The trained model 4 receives the frequency data 33 and the reference waveform data 33A indicating a normal state, and outputs whether the frequency data 33 indicates an abnormal state. This allows the determination system 1 to determine whether one operation corresponding to the frequency data 33 is an abnormal operation, among multiple operations repeatedly performed in the production equipment 2. The trained model 4 also outputs whether the frequency data 33 indicates an abnormal state, based on the frequency data 33 and the reference waveform data 33A. Therefore, the trained model 4 can determine whether the frequency data 33 indicates an abnormal state with high accuracy, compared to when determining whether the frequency data 33 indicates an abnormal state based only on the frequency data 33.

[0039] Here, when the frequency data 33 indicates an abnormal state, the trained model 4 may output a portion of the frequency data 33 that indicates the abnormal state. For example, when the frequency data 33 indicates a normal state, the trained model 4 outputs that fact. Furthermore, when the frequency data 33 indicates an abnormal state, the trained model 4 outputs information indicating which portion of the frequency data 33 indicates the abnormality. This allows the user of the determination system 1 to know the accuracy of the determination result of the determination system 1.

[0040] (4) Configuration of the Learning System The learning system 6 according to the embodiment includes an acquisition unit 11, a conversion unit 12, a memory unit 13, a state acquisition unit 61, and a learning unit 62, as shown in FIG.

[0041] The status acquisition unit 61 acquires data indicating whether a corresponding operation of the production equipment 2 is normal or abnormal for each of the multiple one-dimensional waveform data 31 acquired by the acquisition unit 11 .

[0042] The learning unit 62 generates teacher data 51 and generates a trained model 4 through machine learning based on the teacher data 51.

[0043] The learning system 6 includes a computer system. The computer system is mainly composed of a processor and a memory as hardware. The processor executes a program recorded in the memory of the computer system, thereby realizing the functions of the learning system 6 in the present disclosure.

[0044] Furthermore, it is not essential for the learning system 6 that multiple functions are concentrated in a single housing, and the components of the learning system 6 may be distributed across multiple housings. Furthermore, at least some of the functions of the learning system 6 may be realized by the cloud (cloud computing) or the like.

[0045] Conversely, at least some of the functions of the learning system 6, which are distributed across multiple components in the embodiment, may be integrated into a single housing.

[0046] The teacher data 51 includes a plurality of frequency data 52. Each of the plurality of frequency data 52 is frequency data 33 converted from one-dimensional waveform data 31 acquired in advance. The teacher data 51 further includes a plurality of state data 53. The plurality of state data 53 corresponds one-to-one to the plurality of frequency data 52. Each of the plurality of state data 53 indicates whether a corresponding one of the plurality of frequency data 52 indicates a normal state or an abnormal state. That is, the plurality of frequency data 52 includes a plurality of first frequency data 52A, each of which is known to indicate whether it indicates a normal state or an abnormal state, and one or more second frequency data 52B, each of which indicates a normal state. The plurality of first frequency data 52A includes frequency data indicating an abnormal state and frequency data indicating a normal state. Here, the one or more first frequency data 52A may include part or all of the one or more second frequency data 52B.

[0047] The learning unit 62 generates a trained model 4 by machine learning based on the teacher data 51. Specifically, as shown in Fig. 8 , for a neural network including a first independent block 43a, a second independent block 43b, an integrated block 43c, and a combination layer 44, one or more pieces of first frequency data 52A are input to the first independent block 43a, and one or more pieces of second frequency data 52B are input to the second independent block 43b. Then, the learning unit 62 trains coefficients of spatial filters of the plurality of convolution layers 41 so that an abnormal state is output when the first frequency data 52A is input to the first independent block 43a, and a normal state is output when the first frequency data 52A is input to the first independent block 43a.

[0048] As a result, the trained model 4 learns the relationship between the frequency data 52A indicating an abnormal state and the frequency data 52A indicating a normal state, which are included in the teacher data 51. Specifically, the trained model 4 receives as input one first frequency data 52A and one second frequency data 52B from the plurality of frequency data 52, and learns to output information indicating whether the one first frequency data 52A is in a normal state or an abnormal state. As a result, the trained model 4 estimates whether the frequency data 33 indicates a normal state or an abnormal state based on whether the relationship between the frequency data 33 and the reference waveform data 33A is similar to the relationship between the first frequency data 52A and the second frequency data 52B. In other words, the trained model 4 can estimate whether the one-dimensional waveform data 31 indicates a normal state or an abnormal state, regardless of whether the frequency data 33 is similar to both the first frequency data 52A and the second frequency data 52B or neither.

[0049] Therefore, if the reference waveform data 33A indicates normal operation of the production equipment 2, the judgment unit 14 of the judgment system 1 according to the embodiment can judge whether the frequency data 33 indicates an abnormal state. In other words, it is sufficient that similar changes occur in the one-dimensional waveform data 31 between the normal state and the abnormal state between the production equipment 2 used to generate the trained model 4 and the production equipment 2 that is the target of state judgment by the judgment system 1; the one-dimensional waveform data 31 itself does not need to have the same characteristics. For example, the positional relationship between the mold 21 and the sensor 22 or the shape of the mold 21 may be different between the production equipment 2 for which the judgment system 1 is to judge whether the frequency data 33 indicates an abnormal state and the production equipment 2 from which the teacher data 51 is generated.

[0050] Furthermore, the trained model 4 does not have a fully connected layer. This allows the learning system 6 to reduce the amount of calculations performed by the learning unit 62 while reducing the impact of the trained model 4 on the determination accuracy.

[0051] Here, the plurality of state data 53 may include information indicating a portion indicating an abnormal state when a corresponding one of the plurality of frequency data 52 is in an abnormal state. When the frequency data 33 and the reference waveform data 33A are input, the learning unit 62 generates a trained model 4 that outputs a portion of the frequency data 33 that indicates an abnormal state. This allows the user of the determination system 1 to know the accuracy of the determination result of the determination system 1.

[0052] (5) Effects A determination system 1 according to the embodiment includes an acquisition unit 11, a conversion unit 12, and a determination unit 14. The acquisition unit 11 acquires discontinuous one-dimensional waveform data 31 having a start point and an end point. The conversion unit 12 converts the one-dimensional waveform data 31 into frequency data 33. The determination unit 14 uses a trained model 4, which is a neural network, to determine whether the frequency data 33 indicates an abnormal state based on a combination of the frequency data 33 and reference waveform data 33A. The trained model 4 is generated by machine learning from teacher data 51. The teacher data 51 includes frequency data 52A indicating an abnormal state and frequency data 52B indicating a normal state.

[0053] Moreover, a determination method according to the embodiment is executed by one or more processors. The determination method includes an acquisition step, a conversion step, and a determination step. In the acquisition step, discontinuous one-dimensional waveform data 31 having a start point and an end point is acquired. In the conversion step, the one-dimensional waveform data 31 is converted into frequency data 33. In the determination step, a trained model 4, which is a neural network, is used to determine whether the frequency data 33 indicates an abnormal state based on a combination of the frequency data 33 and reference waveform data 33A. The trained model 4 is generated by machine learning from teacher data 51. The teacher data 51 includes frequency data 52A indicating an abnormal state and frequency data 52A indicating a normal state.

[0054] Furthermore, the program according to the embodiment causes one or more processors to execute the determination method according to the embodiment.

[0055] As a result, the determination system 1 according to the embodiment can generally determine an abnormal state based on the one-dimensional waveform data 31 .

[0056] Furthermore, in the determination system 1 according to the embodiment, the determination unit 14 inputs the frequency data 33 as two-dimensional image data to the trained model 4. This makes it possible for the determination system 1 according to the embodiment to use a machine learning model suitable for image recognition as the trained model 4.

[0057] Furthermore, in the determination system 1 according to the embodiment, the trained model 4 is a convolutional neural network from which the fully connected layer is excluded. The trained model 4 includes a convolutional layer 41 and a pooling layer 42. In the trained model 4, the filter of the convolutional layer 41 is trained on a processing result having a spatial dimension of 1×1. This makes it possible for the determination system 1 according to the embodiment to improve the accuracy of determination by the determination unit 14 while reducing the amount of calculation required to generate the trained model 4.

[0058] Furthermore, in the determination system 1 according to the embodiment, the trained model 4 is a trained model that has learned the relationship between one or more pieces of first frequency data 52A and one or more pieces of second frequency data 52B indicating a normal state, which are included in the teacher data 51. As a result, in the determination system 1 according to the embodiment, as long as the reference waveform data 33A indicates normal operation of the production equipment 2, the determination unit 14 can determine whether the one-dimensional waveform data 31 indicates an abnormal state even if the plurality of frequency data 52 included in the teacher data 51 and the one-dimensional waveform data 31 have different characteristics.

[0059] Furthermore, in the determination system 1 according to the embodiment, when the frequency data 33 indicates an abnormal state, the determination unit 14 may indicate a portion of the frequency data 33 that indicates the abnormal state. As a result, according to the determination system 1 according to the embodiment, the user of the determination system 1 can know the validity of the determination result by the determination unit 14, thereby improving the accuracy of the determination by the determination unit 14.

[0060] The determination system 1 according to the embodiment also includes a storage unit 13 that stores a plurality of frequency data 33. The determination unit 14 sets, as reference waveform data 33A, a representative value of two or more of the plurality of frequency data 33 stored in the storage unit 13. This makes it possible to reduce the influence of variations among the plurality of frequency data 33 on the determination result by the determination unit 14.

[0061] Furthermore, in the determination system 1 according to the embodiment, the one-dimensional waveform data 31 indicates one operation among a plurality of operations repeatedly performed in the production facility 2. This enables the determination system 1 to determine whether or not an abnormality has occurred in each of the plurality of operations repeatedly performed in the production facility 2.

[0062] Furthermore, in the determination system 1 according to the embodiment, the acquisition unit 11 extracts one-dimensional waveform data 31 from data 32 acquired from a sensor 22 installed in the production facility 2. This makes it easy for the determination system 1 to determine whether or not an abnormality has occurred in each of a plurality of operations repeatedly performed in the production facility 2.

[0063] Furthermore, in the determination system 1 according to the embodiment, the production facility 2 is a molding machine that uses a mold 21. The abnormal state determined by the determination unit 14 is a slug defect in the mold 21. This enables the determination system 1 to detect slugs in the production facility 2 at an early stage and reduce the occurrence of consecutive manufacturing defects.

[0064] Moreover, the learning system 6 according to the embodiment includes an acquisition unit 11, a conversion unit 12, a state acquisition unit 61, and a learning unit 62. The acquisition unit 11 acquires a plurality of discontinuous one-dimensional waveform data 31 each having a start point and an end point. The conversion unit 12 converts the plurality of one-dimensional waveform data 31 into a plurality of frequency data 33. The state acquisition unit 61 acquires a plurality of state data 53 indicating whether each of the plurality of frequency data 33 indicates an abnormal state. The learning unit 62 generates a trained model 4, which is a neural network, using training data including the plurality of frequency data 33. The trained model 4 receives as input first frequency data 52A, which is one of the plurality of frequency data 33, and second frequency data 52B, which is one of the plurality of frequency data 33 and does not indicate an abnormal state, and outputs whether the first frequency data 52A indicates an abnormal state.

[0065] A learning method according to the embodiment is executed by one or more processors. The learning method includes an acquisition step, a conversion step, a state acquisition step, and a learning step. In the acquisition step, a plurality of discontinuous one-dimensional waveform data 31 each having a start point and an end point is acquired. In the conversion step, the plurality of one-dimensional waveform data 31 is converted into a plurality of frequency data 33. In the state acquisition step, a plurality of state data 53 indicating whether each of the plurality of frequency data 33 indicates an abnormal state is acquired. In the learning step, a trained model 4, which is a neural network, is generated using teacher data 51. The teacher data 51 includes the plurality of frequency data 33. The trained model 4 receives as input first frequency data 52A, which is one of the plurality of frequency data 33, and second frequency data 52B, which is one of the plurality of frequency data 33 and indicates a normal state, and outputs whether the first frequency data 52A indicates an abnormal state.

[0066] Furthermore, the program according to the embodiment causes one or more processors to execute the learning method according to the embodiment.

[0067] This allows the learning system 6 to determine abnormal conditions in a general manner based on the one-dimensional waveform data 31.

[0068] (Other Modifications According to the Embodiment) (1) Although the learning system 6 according to the embodiment is a system independent of the determination system 1, the learning system 6 and the determination system 1 may share some of the configuration. For example, the learning system 6 and the determination system 1 may share the acquisition unit 11, the conversion unit 12, and the storage unit 13. Furthermore, for example, the learning unit 62 of the learning system 6 may output the trained model 4 to the determination unit 14 of the determination system 1.

[0069] (Aspect) A determination system (1) according to a first aspect includes an acquisition unit (11), a conversion unit (12), and a determination unit (14). The acquisition unit (11) acquires discontinuous one-dimensional waveform data having a start point and an end point. The conversion unit (12) converts the one-dimensional waveform data (31) into frequency data (33). The determination unit (14) uses a trained model (4) that is a neural network to determine whether the frequency data (33) indicates an abnormal state based on a combination of the frequency data (33) and reference waveform data (33A). The trained model (4) is generated by machine learning from teacher data (51). The teacher data (51) includes frequency data (52A) corresponding to an abnormal state and frequency data (52B) corresponding to a normal state.

[0070] According to the determination system (1) of the above aspect, it becomes possible to determine an abnormal state in a general manner based on one-dimensional waveform data (31).

[0071] In the determination system (1) according to the second aspect, in the first aspect, the determination unit (14) inputs frequency data (33) as two-dimensional image data to the trained model (4).

[0072] According to the determination system (1) relating to the above aspect, it is possible to use a machine learning model suitable for image recognition as the trained model (4).

[0073] In the determination system (1) according to the third aspect, in the second aspect, the trained model (4) is a convolutional neural network in which the fully connected layer is excluded. The trained model (4) includes a convolutional layer (41) and a pooling layer (42). In the trained model (4), the filter of the convolutional layer (41) is trained on a processing result having a spatial dimension of 1×1.

[0074] According to the determination system (1) relating to the above aspect, it is possible to improve the accuracy of determination by the determination unit (14) while reducing the amount of calculation required to generate the trained model (4).

[0075] In the judgment system (1) according to the fourth aspect, in the second or third aspect, the trained model (4) is a trained model that has learned the relationship between frequency data (52A) indicating an abnormal state and frequency data (52B) indicating a normal state, which are included in the teacher data (51).

[0076] According to the determination system (1) relating to the above aspect, even if the plurality of frequency data (52) and the one-dimensional waveform data (31) have different characteristics, the determination unit (14) can determine whether or not the one-dimensional waveform data (31) indicates an abnormal state.

[0077] In the determination system (1) according to the fifth aspect, in any of the second to fourth aspects, the determination unit (14) indicates a portion of the frequency data (33) that indicates an abnormal state when the frequency data (33) indicates an abnormal state.

[0078] According to the determination system (1) of the above aspect, the user of the determination system (1) can know the validity of the determination result by the determination unit (14), thereby improving the accuracy of the determination by the determination unit (14).

[0079] A determination system (1) according to a sixth aspect is any one of the first to fifth aspects, further comprising a storage unit (13). The storage unit (13) stores a plurality of frequency data (33). The determination unit (14) determines a representative value of two or more of the plurality of frequency data (33) stored in the storage unit (13) as reference waveform data (33A).

[0080] According to the determination system (1) of the above aspect, it is possible to reduce the influence of variations among the plurality of frequency data (33) on the determination result by the determination unit (14).

[0081] In the determination system (1) according to the seventh aspect, in any of the first to sixth aspects, the one-dimensional waveform data (31) indicates one operation among a plurality of operations repeatedly performed in the production equipment (2).

[0082] According to the determination system (1) of the above aspect, it is possible to determine whether or not an abnormality has occurred in each of a plurality of operations that are repeatedly performed in the production equipment (2).

[0083] In the determination system (1) according to the eighth aspect, in the seventh aspect, the acquisition unit (11) extracts one-dimensional waveform data (31) from data (32) acquired from a sensor (22) installed in the production facility (2).

[0084] According to the determination system (1) of the above aspect, it becomes easy to determine whether or not an abnormality has occurred in each of a plurality of operations that are repeatedly performed in the production equipment (2).

[0085] In the judgment system (1) according to the ninth aspect, in the seventh or eighth aspect, the production equipment (2) is a molding machine using a mold (21). The abnormal state judged by the judgment unit (14) is a slug lift defect in the mold (21).

[0086] According to the determination system (1) of the above aspect, it is possible to detect scraps in the production equipment (2) at an early stage and reduce the occurrence of consecutive manufacturing defects.

[0087] A learning system (6) according to a tenth aspect includes an acquisition unit (11), a conversion unit (12), a state acquisition unit (61), and a learning unit (62). The acquisition unit (11) acquires a plurality of discontinuous one-dimensional waveform data (31) having a start point and an end point. The conversion unit (12) converts the plurality of one-dimensional waveform data (31) into a plurality of frequency data (33). The state acquisition unit (61) acquires a plurality of state data (53) indicating whether each of the plurality of frequency data (33) indicates an abnormal state. The learning unit (62) generates a trained model (4), which is a neural network, using training data including the plurality of frequency data (33). The trained model (4) receives as input first frequency data (52A), which is one of the plurality of frequency data (33), and second frequency data (52B), which is one of the plurality of frequency data (33) and indicates a normal state, and outputs whether or not the first frequency data (52A) indicates an abnormal state.

[0088] According to the learning system (6) of the above aspect, it becomes possible to determine an abnormal state in a general manner based on one-dimensional waveform data (31).

[0089] A determination method according to an eleventh aspect is executed by one or more processors. The determination method includes an acquisition step, a conversion step, and a determination step. In the acquisition step, discontinuous one-dimensional waveform data (31) having a start point and an end point is acquired. In the conversion step, the one-dimensional waveform data (31) is converted into frequency data (33). In the determination step, a trained model (4) that is a neural network is used to determine whether or not the frequency data (33) indicates an abnormal state based on a combination of the frequency data (33) and reference waveform data (33A). The trained model (4) is generated by machine learning from teacher data (51). The teacher data (51) includes frequency data (52A) corresponding to an abnormal state and frequency data (52A) corresponding to a normal state.

[0090] According to the determination method of the above aspect, it becomes possible to determine an abnormal state in a general manner based on one-dimensional waveform data (31).

[0091] A program according to a twelfth aspect causes one or more processors to execute the determination method according to the eleventh aspect.

[0092] According to the program of the above aspect, it becomes possible to generally determine an abnormal state based on one-dimensional waveform data (31).

[0093] A learning method according to a thirteenth aspect is executed by one or more processors. The learning method includes an acquisition step, a conversion step, a state acquisition step, and a learning step. In the acquisition step, a plurality of discontinuous one-dimensional waveform data (31) having a start point and an end point is acquired. In the conversion step, the plurality of one-dimensional waveform data (31) is converted into a plurality of frequency data (33). In the state acquisition step, a plurality of state data (53) indicating whether each of the plurality of frequency data (33) indicates an abnormal state is acquired. In the learning step, a trained model (4) that is a neural network is generated using teacher data (51) including the plurality of frequency data (33). The trained model (4) receives as input first frequency data (52A) that is one of the plurality of frequency data (33) and second frequency data (52B) that is one of the plurality of frequency data (33) and indicates a normal state, and outputs whether the first frequency data (52A) indicates an abnormal state.

[0094] According to the learning method of the above aspect, it becomes possible to determine an abnormal state in a general manner based on one-dimensional waveform data (31).

[0095] A program according to a fourteenth aspect causes one or more processors to execute the learning method according to the thirteenth aspect.

[0096] According to the program of the above aspect, it becomes possible to generally determine an abnormal state based on one-dimensional waveform data (31).

[0097] REFERENCE SIGNS LIST 1 Determination system 2 Production equipment 21 Mold 22 Sensor 11 Acquisition unit 12 Conversion unit 13 Memory unit 14 Determination unit 31 One-dimensional waveform data 32 Sensor data (data) 33 Frequency data 33A Reference waveform data 4 Trained model 41 Convolution layer 42 Pooling layer 51 Teacher data 52 Frequency data 52A First frequency data (frequency data) 52B Second frequency data (frequency data)

Claims

1. A determination system comprising: an acquisition unit that acquires discontinuous one-dimensional waveform data having a start point and an end point; a conversion unit that converts the one-dimensional waveform data into frequency data; and a determination unit that determines whether the frequency data indicates an abnormal state from a combination of the frequency data and reference waveform data, using a trained model that is a neural network generated by machine learning from teacher data including frequency data indicating an abnormal state and frequency data indicating a normal state.

2. The determination system according to claim 1, wherein the determination unit inputs the frequency data to the trained model as two-dimensional image data.

3. The determination system according to claim 2, wherein the trained model is a convolutional neural network in which a fully connected layer is excluded, and includes a convolutional layer and a pooling layer, and a filter of the convolutional layer is trained on a processing result having a spatial dimension of 1×1.

4. The judgment system according to claim 2 or 3, wherein the trained model is a trained model that has learned the relationship between frequency data indicating an abnormal state and frequency data indicating a normal state, which are included in the training data.

5. The determination system according to any one of claims 2 to 4, wherein, when the frequency data indicates an abnormal state, the determination unit indicates a portion of the frequency data that indicates an abnormal state.

6. A judgment system according to any one of claims 1 to 5, further comprising a memory unit that stores a plurality of said frequency data, and wherein the judgment unit uses a representative value of two or more of said plurality of frequency data stored in the memory unit as said reference waveform data.

7. The determination system according to any one of claims 1 to 6, wherein the one-dimensional waveform data represents one operation among a plurality of operations repeatedly performed in a production facility.

8. The determination system according to claim 7, wherein the acquisition unit extracts the one-dimensional waveform data from data acquired from a sensor installed in the production facility.

9. The judgment system according to claim 7 or 8, wherein the production equipment is a molding machine that uses a mold, and the abnormal state judged by the judgment unit is a slug lift defect in the mold.

10. A learning system comprising: an acquisition unit that acquires a plurality of discontinuous one-dimensional waveform data each having a start point and an end point; a conversion unit that converts the plurality of one-dimensional waveform data into a plurality of frequency data; a state acquisition unit that acquires a plurality of state data indicating whether each of the plurality of frequency data indicates an abnormal state; and a learning unit that generates a trained model that is a neural network using teacher data including the plurality of frequency data, wherein the trained model receives as input first frequency data that is one of the plurality of frequency data and second frequency data that is one of the plurality of frequency data and does not indicate an abnormal state, and outputs whether the first frequency data indicates an abnormal state.

11. A determination method executed by one or more processors, comprising: an acquisition step of acquiring discontinuous one-dimensional waveform data having a start point and an end point; a conversion step of converting the one-dimensional waveform data into frequency data; and a determination step of determining whether the frequency data indicates an abnormal state from a combination of the frequency data and reference waveform data, using a trained model that is a neural network generated by machine learning from teacher data including frequency data indicating an abnormal state and frequency data indicating a normal state.

12. A program causing one or more processors to execute the determination method according to claim 11.

13. A learning method executed by one or more processors, comprising: an acquisition step of acquiring a plurality of discontinuous one-dimensional waveform data each having a start point and an end point; a conversion step of converting the plurality of one-dimensional waveform data into a plurality of frequency data; a state acquisition step of acquiring a plurality of state data indicating whether each of the plurality of frequency data indicates an abnormal state; and a learning step of generating a trained model which is a neural network using training data including the plurality of frequency data, wherein the trained model receives as input first frequency data which is one of the plurality of frequency data and second frequency data which is one of the plurality of frequency data and indicates a normal state, and outputs whether the first frequency data indicates an abnormal state.

14. A program causing one or more processors to execute the learning method of claim 13.

Images