Defect determination device, defect classification device, model generation device, defect determination method, defect classification method, model generation method, learned model, and program

Abstract

To provide a defect determination device which can determine whether or not tightening of a screw is completed normally regardless of a type of a defect, and to provide a defect classification device, a model generation device, a defect determination method, a defect classification method, a model generation method, a learned model, and a program.SOLUTION: A defect determination device according to an aspect of the disclosure includes: a physical quantity acquisition unit which acquires a target physical quantity being a physical quantity generated in a power source during a new screw tightening; a model acquisition unit which acquires a first learned model obtained through non-defective product learning of a physical quantity generated in the power source during a past screw tightening that has been normally completed; and a determination unit which determines whether or not the new screw tightening has been normally completed by applying the target physical quantity to the first learned model.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The present disclosure relates to a defect determination device, a defect classification device, a model generation device, a defect determination method, a defect classification method, a model generation method, a trained model, and a program. [Background technology]

[0002] In existing automatic screw tightening devices that automatically tighten screws, a technique is known that automatically determines whether a screw has been tightened properly by threshold judgment using physical quantities such as the torque and rotational speed transmitted from the motor to the driver.

[0003] For example, Patent Document 1 discloses a technology for estimating the axial force of a screw generated by tightening a screw and comparing it with a threshold value to determine whether the screw is defective. The technology disclosed in Patent Document 1 can detect thread defects, where the threads are stripped, and bottoming, where the screw tip contacts the bottom of the screw hole before the seating surface of the screw head contacts the workpiece. However, it is difficult to detect jamming, where dirt or other foreign matter is trapped when a screw is tightened into a workpiece, because the axial force generated is the same as that generated when a screw is tightened normally.

[0004] Furthermore, for example, Patent Document 2 discloses a technology for determining defects by comparing the amount of change in the axial position of the screwdriver from the temporary seating stage to the final tightening process with a threshold value. The technology disclosed in Patent Document 2 is capable of detecting bottoming out and foreign matter jamming as described above. However, it is difficult to detect screw float, which occurs when a screw is tightened into a workpiece while inserted at an angle relative to the screw hole, because the amount of change in the axial position of the screwdriver is about the same as in normal screw tightening.

[0005] For example, Patent Document 3 discloses a technology for detecting defects by extracting feature values ​​from torque and rotational speed waveforms and comparing them with thresholds. Patent Document 3 also discloses selecting multiple feature values ​​that contribute to improving accuracy, extracting these selected feature values ​​from torque and rotational speed waveforms, and integrating them into a single numerical index. Patent Document 3 also discloses using machine learning to select feature values ​​that are highly correlated with conforming and non-conforming screw tightening states and have small variations, in order to reduce the burden of selecting multiple feature values ​​that contribute to improving accuracy. The technology disclosed in Patent Document 3 can detect the above-mentioned thread defects and foreign object jamming. However, because the technology disclosed in Patent Document 3 sets a dedicated threshold for each type of defect, it is difficult to detect types of defects for which no threshold is set, such as screw loosening. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent No. 7375684 [Patent Document 2] Patent No. 7070467 [Patent Document 3] Patent No. 7186905 Summary of the Invention [Problem to be solved by the invention]

[0007] Each of the above-mentioned existing technologies is based on a design concept that targets detecting a specific type of defect. Therefore, it is difficult for each of the above-mentioned existing technologies to deal with defects outside the design concept (defects other than the specific type). Therefore, it is difficult for each of the above-mentioned existing technologies to determine whether screw tightening has been completed successfully, regardless of whether the defect is within the design concept or outside the design concept, for example, regardless of the type of defect.

[0008] The present disclosure has been made in consideration of the above circumstances, and contributes to providing a defect determination device, a defect classification device, a model generation device, a defect determination method, a defect classification method, a model generation method, a trained model, and a program that are capable of determining whether screw tightening has been completed successfully regardless of the type of defect. [Means for solving the problem]

[0009] One aspect of the defect judgment device of the present disclosure includes a physical quantity acquisition unit that acquires a target physical quantity, which is a physical quantity that occurs in a power source when a new screw is tightened; a model acquisition unit that acquires a first trained model that has been trained as a good product based on the physical quantity that occurred in the power source when a previous screw was tightened successfully; and a judgment unit that applies the target physical quantity to the first trained model and judges whether the new screw tightening has been completed successfully.

[0010] One aspect of the defect classification device of the present disclosure includes a physical quantity acquisition unit that acquires a target physical quantity, which is a physical quantity that occurs in a power source during new screw tightening; a model acquisition unit that acquires a trained model that uses as input the physical quantity that occurred in the power source during a previous screw tightening in which an abnormality occurred, and outputs a screw tightening defect classification result; and a classification unit that applies the target physical quantity to the trained model to classify the new screw tightening defect.

[0011] One aspect of the model generation device of the present disclosure includes a physical quantity acquisition unit that acquires a physical quantity generated in a power source when a screw is tightened successfully, and a learning unit that performs quality learning on the physical quantity and generates a trained model that uses as input the physical quantity generated in the power source when a new screw is tightened and outputs whether the new screw tightening was completed successfully.

[0012] One aspect of the model generation device disclosed herein includes a physical quantity acquisition unit that acquires physical quantities that occur in a power source when an abnormality occurs in a screw tightening, and a learning unit that learns learning data including the physical quantities and generates a trained model that uses as input the physical quantities that occur in the power source when a new screw is tightened and outputs a fault classification result for the new screw tightening.

[0013] One aspect of the defect determination method of the present disclosure includes a physical quantity acquisition step in which a physical quantity acquisition unit acquires a target physical quantity, which is a physical quantity that occurs in a power source when a new screw is tightened; a model acquisition step in which a model acquisition unit acquires a first trained model that has been trained as a good product using the physical quantity that occurred in the power source when a previous screw was tightened successfully; and a determination step in which a determination unit applies the target physical quantity to the first trained model to determine whether the new screw tightening was completed successfully.

[0014] One aspect of the defect classification method of the present disclosure includes a physical quantity acquisition step in which a physical quantity acquisition unit acquires a target physical quantity, which is a physical quantity that occurred in a power source during new screw tightening; a model acquisition step in which a model acquisition unit acquires a trained model that uses as input the physical quantity that occurred in the power source during a previous screw tightening in which an abnormality occurred, and as output a screw tightening defect classification result; and a classification step in which a classification unit applies the target physical quantity to the trained model to classify the new screw tightening defect.

[0015] One aspect of the model generation method of the present disclosure includes a physical quantity acquisition step in which a physical quantity acquisition unit acquires a physical quantity that occurs in a power source when a screw is tightened successfully, and a learning step in which a learning unit performs quality learning on the physical quantity and generates a trained model that uses as input the physical quantity that occurs in the power source when a new screw is tightened and outputs whether the new screw tightening was completed successfully.

[0016] One aspect of the model generation method of the present disclosure includes a physical quantity acquisition step in which a physical quantity acquisition unit acquires a physical quantity that occurs in a power source when a screw is tightened and an abnormality occurs, and a learning step in which a learning unit learns learning data including the physical quantity and generates a trained model that uses as input the physical quantity that occurs in the power source when a new screw is tightened and outputs a fault classification result for the new screw tightening.

[0017] One aspect of the trained model of the present disclosure is generated by the above-mentioned model generation device or the above-mentioned model generation method.

[0018] One aspect of the program of the present disclosure is for causing a computer to execute the above method. [Effects of the Invention]

[0019] According to the present disclosure, it is possible to provide a defect determination device, a defect classification device, a model generation device, a defect determination method, a defect classification method, a model generation method, a trained model, and a program that can determine whether screw tightening has been completed successfully regardless of the type of defect. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is a schematic diagram showing an example of the configuration of a screw fastening device according to a first embodiment. [Figure 2] FIG. 2 is a diagram showing an example of a waveform of torque generated in a motor from the start to the end of screw tightening by the screw tightening device of the first embodiment. [Figure 3] FIG. 3 is a diagram showing a comparison example between a torque waveform in a normal system and a torque waveform in an abnormal system in the screw fastening device of the first embodiment. [Figure 4] FIG. 4 is a block diagram showing an example of the hardware configuration of the defect determination device according to the first embodiment. [Figure 5] FIG. 5 is a block diagram illustrating an example of the functional configuration of the defect determination device according to the first embodiment. [Figure 6] FIG. 6 is an explanatory diagram illustrating an example of a non-defective product learning method performed by the learning unit of the first embodiment. [Figure 7] FIG. 7 is an explanatory diagram illustrating an example of a non-defective product learning method performed by the learning unit of the first embodiment. [Figure 8] FIG. 8 is an explanatory diagram illustrating an example of a defect determination method using the first trained model by the determination unit of the first embodiment. [Figure 9] FIG. 9 is a flowchart showing an example of a screw tightening process performed by the screw tightening device of the first embodiment. [Figure 10] FIG. 10 is a flowchart showing an example of a learning process performed by the defect determination device of the first embodiment. [Figure 11] FIG. 11 is a flowchart showing an example of a defect determination process performed by the defect determination device of the first embodiment. [Figure 12] FIG. 12 is a block diagram illustrating an example of the functional configuration of a defect determination device according to the second embodiment. [Figure 13] FIG. 13 is an explanatory diagram illustrating an example of a learning method for the second trained model by the learning unit of the second embodiment. [Figure 14] FIG. 14 is a flowchart showing an example of a defect classification process performed by the defect determination device of the second embodiment. [Figure 15] FIG. 15 is a block diagram illustrating an example of the functional configuration of a defect determination device according to the third embodiment. [Figure 16] FIG. 16 is an explanatory diagram showing an example of a preprocessing technique for a target physical quantity by the preprocessing unit according to the third embodiment. [Figure 17] FIG. 17 is a flowchart showing an example of advance preparation processing for performing preprocessing by the defect determination device of the third embodiment. [Figure 18] FIG. 18 is a flowchart showing an example of pre-processing by the defect determination device 300 of the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0021] Hereinafter, embodiments of the present disclosure (hereinafter simply referred to as "present embodiments") will be described in detail with reference to the drawings. Note that the present disclosure is not limited to the following embodiments. Furthermore, the following embodiments and modifications can be combined as appropriate.

[0022] (First embodiment) First, an overview of the screw tightening device will be described.

[0023] 1 is a schematic diagram showing an example of the configuration of a screw fastening device 10 according to a first embodiment. The screw fastening device 10 is a device that automatically fastens a screw 5 into a screw hole 2 formed in a workpiece 1. The screw hole 2 corresponds to a female thread, and the screw 5 corresponds to a male thread.

[0024] As shown in FIG. 1, the screw tightening device 10 includes a driver bit 20, a motor 30, a stage 40, a motor drive device 50, a stage drive device 60, a main control device 70, and a defect determination device 100.

[0025] The driver bit 20 is a rotary tool used to tighten the screw 5 into the screw hole 2. The driver bit 20 is attached to a rotary shaft (not shown) of the motor 30, and rotates around its own axis in response to the rotation of the rotary shaft. The driver bit 20 rotates around its axis while holding the screw 5 by suction at its tip, thereby transmitting the torque generated in the motor 30 to the screw 5 and fastening the screw 5 into the screw hole 2.

[0026] The motor 30 is a power source for screw tightening. Examples of the motor 30 include, but are not limited to, a servo motor. The motor 30 generates torque for fastening the screw 5 into the screw hole 2 by rotating its rotary shaft in accordance with a drive signal output from the motor drive device 50. Examples of the drive signal include, but are not limited to, a PWM (Pulse Width Modulation) signal. The motor 30 also includes an encoder (not shown), which detects the torque and rotation angle of the rotary shaft of the motor 30 as analog values ​​and outputs them to the motor drive device 50 as feedback signals.

[0027] The stage 40 is movable in the up / down, left / right, and depth directions, and is connected to a motor 30 and a driver bit 20 via the motor 30. The stage 40 moves in accordance with a drive signal output from a stage drive device 60, thereby moving the driver bit 20 to a desired position. The stage 40 is equipped with a motor (not shown), and moves by rotating the motor in accordance with a drive signal output from the stage drive device 60. The motor (not shown) of the stage 40 is equipped with an encoder (not shown), and the detection result of this encoder is output to the stage drive device 60 as a feedback signal. In the first embodiment, an example is described in which the stage 40 is movable in three axes (up / down, left / right, and depth), but this is not limited thereto. For example, if the workpiece 1 can be moved in the left / right and depth directions, the stage 40 may be movable in one axis (up / down).

[0028] The motor drive unit 50 drives the rotation shaft of the motor 30 to rotate. Examples of the motor drive unit 50 include, but are not limited to, drive circuits such as a motor driver and a motor amplifier. The motor drive unit 50 receives a motor control signal from the main control unit 70 that indicates the rotation angle, rotation speed, torque, etc. of the motor 30. The motor drive unit 50 generates a drive signal that drives the rotation shaft of the motor 30 to rotate according to the control content indicated by the received motor control signal, and outputs the drive signal to the motor 30.

[0029] The motor driver 50 also receives the above-mentioned feedback signal from the motor 30. Based on the received feedback signal, the motor driver 50 generates a waveform of the torque actually generated in the motor 30 and a waveform of the actual rotational speed of the motor 30, and transmits these to the main control device 70. For example, the motor driver 50 samples the received feedback signal at a fixed sample interval (e.g., 2 ms) and quantizes it to generate waveform data.

[0030] The stage driving device 60 moves the stage 40. Examples of the stage driving device 60 include, but are not limited to, driving circuits such as a motor driver and a motor amplifier. The stage driving device 60 receives a stage control signal from the main control device 70 that instructs the movement position of the stage 40, generates a drive signal that moves the stage 40 to the instructed movement position, and outputs it to the stage 40. The stage driving device 60 also receives the above-mentioned feedback signal from the stage 40 and sends it to the main control device 70.

[0031] The main control device 70 performs overall control of the screw fastening device 10. The main control device 70 may be realized, for example, by a PLC (Programmable Logic Controller), or may be realized as software using a general-purpose processor or the like. The main control device 70 generates the motor control signals and stage control signals described above in accordance with a pre-registered control flow. The main control device 70 controls screw fastening by the driver bit 20 by outputting the generated motor control signals to the motor drive device 50. The main control device 70 also controls the movement of the driver bit 20 by outputting the generated stage control signals to the stage drive device 60. By controlling the driver bit 20 in this manner, the main control device 70 automates screw fastening by the screw fastening device 10.

[0032] The main control device 70 also outputs data on the torque waveform and rotational speed waveform of the motor 30 received from the motor drive device 50 to the defect determination device 100, and receives defect determination results using these waveform data from the defect determination device 100.

[0033] The defect determination device 100 determines whether or not the screw tightening by the screw tightening device 10 has been completed normally, regardless of the type of defect. The defect determination device 100 may be realized as software using a general-purpose processor or the like, or may be realized by a PLC. Furthermore, the defect determination device 100 may be realized as the same device as the main control device 70, or may be realized as a device separate from the main control device 70.

[0034] The defect determination device 100 uses the waveforms of at least one of the torque and rotational speed of the motor 30 received from the main control device 70 to determine whether the screw tightening by the screw tightening device 10 was completed normally without any difference from the characteristics of normal screw tightening.

[0035] Specifically, the defect determination device 100 generates a trained model by previously learning the waveforms of at least one of the torque and rotational speed generated in the motor 30 during past screw tightening operations that were completed successfully. Note that the trained model is unsupervised machine learning that learns training data of non-defective products to generate a trained model. The trained model generated by the non-defective product learning can determine whether the object to be identified is a non-defective product or a defective product. The defect determination device 100 applies the received waveforms of at least one of the torque and rotational speed of the motor 30 to the generated trained model to determine whether the screw tightening by the screw tightening device 10 was completed successfully without any discrepancy with the characteristics of normal screw tightening. The defect determination device 100 transmits the defect determination result to the main control device 70.

[0036] Next, the waveform of the torque generated in the motor 30 will be described.

[0037] Fig. 2 is a diagram showing an example of the waveform of torque generated in the motor 30 from the start to the end of screw tightening by the screw tightening device 10 of the first embodiment. In the example shown in Fig. 2, the vertical axis represents torque and the horizontal axis represents time. The torque waveform shown in Fig. 2 is a normal waveform in which screw tightening is completed normally without any defects.

[0038] As shown in FIG. 2, screw tightening by the screw tightening device 10 is comprised of five steps: alignment, screwing, temporary seating, final tightening, and torque maintenance.

[0039] Alignment is a process of moving the driver bit 20 so that the screw 5 held by suction on the driver bit 20 is positioned directly above the screw hole 2. In the alignment process, the screw 5 is held by suction on the driver bit 20 and is not tightened into the screw hole 2. For this reason, the torque is maintained at a relatively low value.

[0040] The screwing process is a process in which the driver bit 20 is rotated while being moved straight down, and the screw 5 held by suction is screwed into the screw hole 2. During the screwing process, the screw 5 is screwed into the female-threaded screw hole 2, and is not tightened into the screw hole 2. For this reason, the torque is maintained at a relatively low value.

[0041] Temporary seating is a state in which the bearing surface of the screw head of the screw 5 is in contact with the workpiece 1. Temporary seating is the stage in which the bearing surface of the screw 5 comes into contact with the workpiece 1 and the screw 5 begins to be tightened into the screw hole 2. For this reason, the torque starts to rise from a relatively low value.

[0042] Final tightening is the process of further tightening the screw 5 from the provisionally seated state. In the final tightening process, when the screw 5 is tightened into the screw hole 2, the friction that occurs between the seating surface of the screw 5 and the workpiece 1 and between the screw 5 and the screw hole 2 increases. For this reason, a torque that exceeds this friction is used, and the torque value increases rapidly.

[0043] Torque maintenance is a process in which the torque at the time of final tightening is maintained for a certain period of time after the completion of final tightening. For this reason, the torque is maintained at a relatively high value. Although not shown in the figure, once the torque maintenance is completed, the torque is released and the torque drops suddenly.

[0044] In this way, the torque waveform during screw tightening by the screw tightening device 10 reflects the characteristics of each screw tightening process. Therefore, when a screw is tightened by the screw tightening device 10, there is an ideal torque waveform for the screw tightening flow. Note that the time and torque required for screw tightening differ depending on the size of the screw, such as its length and thickness, and therefore an ideal torque waveform exists for each screw size. Therefore, the torque waveform when a screw of a certain size is successfully tightened will be a normal torque waveform that is close to the ideal torque waveform for that size. On the other hand, when a screw is tightened due to some kind of defect, the effect of the defect will be reflected in the torque waveform, resulting in an abnormal torque waveform that deviates from the normal torque waveform.

[0045] Fig. 3 is a diagram showing a comparison example between a normal torque waveform and an abnormal torque waveform using the screw fastening device 10 of the first embodiment. The torque waveform data shown in Fig. 3 are all waveforms obtained when screws of the same size are fastened using the screw fastening device 10. As with the example shown in Fig. 2, the vertical axis represents torque and the horizontal axis represents time.

[0046] In the example shown in Figure 3, the group of waveforms grouped by reference numeral 51 corresponds to normal torque waveforms, and the two waveforms grouped by reference numeral 53 correspond to abnormal torque waveforms. Note that all of the abnormal torque waveforms are waveforms that occur when a foreign object has been trapped, which is a defect. Foreign object trapping is a defect in which a screw is fastened to a fastened object with debris or other particles trapped between it and the workpiece. In the case of foreign object trapping, debris or other particles are trapped between the seating surface of the screw head and the workpiece, so the time until provisional seating is achieved is shorter than when the screw is tightened normally. For this reason, as shown in Figure 3, when foreign object trapping occurs, the torque waveform rises faster than in a normal system, resulting in an abnormal torque waveform that deviates from the normal torque waveform.

[0047] In this way, the torque waveform of an abnormal system tends to deviate from the torque waveform of a normal system. For this reason, the defect determination device 100 of the first embodiment generates a trained model that has been trained on the torque waveform of a normal system as a non-defective product, and uses the generated trained model to determine whether the screw tightening by the screw tightening device 10 has been completed normally without any discrepancies with the characteristics of normal screw tightening.

[0048] Next, the configuration of the defect determination device 100 will be described.

[0049] 4 is a block diagram showing an example of the hardware configuration of the defect determination device 100 of the first embodiment. As shown in FIG. 4, the defect determination device 100 includes a processor 101, a memory 103, an auxiliary storage device 105, an input / output interface 107, a communication interface 109, and various buses 111.

[0050] The processor 101, memory 103, auxiliary storage device 105, input / output interface 107, and communication interface 109 are connected via various buses 111. In this way, the first embodiment will be described taking as an example a case where the defect determination device 100 has an existing hardware configuration using an existing computer, but is not limited to this. As described above, the defect determination device 100 is realized by a PLC, and may also have a dedicated hardware configuration.

[0051] The processor 101 controls the overall operation of the defect determination device 100. The processor 101 may be, for example, a CPU (Central Processing Unit), but is not limited to this. The number of CPUs may be one or more, and may be single-core or multi-core.

[0052] Examples of the memory 103 include, but are not limited to, a ROM (Read Only Memory) and a RAM (Random Access Memory). The ROM stores various programs such as a program for controlling the defect determination device 100 and a machine learning program. The RAM is used as a working area when the processor 101 performs various controls based on the programs stored in the ROM.

[0053] The auxiliary storage device 105 stores the various programs and data for machine learning described above. The various programs described above may be stored in at least one of the memory 103 and the auxiliary storage device 105. Examples of the auxiliary storage device 105 include, but are not limited to, at least one of existing storage devices capable of magnetically, electrically, or optically storing data, such as a hard disk drive (HDD), a solid state drive (SSD), and a digital versatile disc (DVD). The auxiliary storage device 105 may be built into the defect determination device 100 or may be externally attached to the defect determination device 100 via an interface such as a universal serial bus (USB). The auxiliary storage device 105 may also be a network-attached storage (NAS) connected via a network such as a local area network (LAN) or a wide area network (WAN).

[0054] The input / output interface 107 is an interface between the defect determination device 100 and various input devices and display devices used during machine learning and defect determination. Examples of the various input devices include, but are not limited to, a keyboard, a mouse, and a touch panel. Examples of the various display devices include, but are not limited to, a liquid crystal display, an organic electroluminescence (EL) display, and a touch panel display.

[0055] Examples of the communication interface 109 include, but are not limited to, a communication interface for a wired LAN and a wireless communication interface for a wireless LAN. The communication interface 109 may be used to acquire the above-mentioned various programs and data for machine learning from the outside, or may be used to output the defect determination result by the defect determination device 100 to the outside.

[0056] Fig. 5 is a block diagram showing an example of the functional configuration of the defect determination device 100 according to the first embodiment. As shown in Fig. 5, the defect determination device 100 includes a physical quantity acquisition unit 121, a learning unit 123, a model acquisition unit 125, and a determination unit 127. The physical quantity acquisition unit 121, the learning unit 123, the model acquisition unit 125, and the determination unit 127 can be realized by, for example, the processor 101 and the memory 103 described in Fig. 4.

[0057] For example, the processor 101 reads a machine learning program stored in the memory 103 (ROM) or the auxiliary storage device 105, or obtained from an external device via the communication interface 109 over a network, and loads the program into the memory 103 (RAM). The processor 101 executes various processes in accordance with the loaded program, thereby realizing each of the above-mentioned functional units as software.

[0058] First, a description will be given of the functions of the physical quantity acquisition unit 121 and the learning unit 123 during non-defective product learning performed in the defect determination device 100. Note that non-defective product learning is performed in advance prior to the defect determination described later.

[0059] The physical quantity acquisition unit 121 acquires the physical quantity generated in the power source when a screw is tightened successfully. The physical quantity acquisition unit 121 acquires the physical quantities of n (n≧2) screw tightenings that have been completed successfully. It is assumed that the same size screws are used in all n screw tightenings. The physical quantity is at least one of the waveform of the torque and the waveform of the rotation speed of the power source. In the first embodiment, the power source is the motor 30 and the physical quantity is the waveform of the torque of the motor 30, but the present invention is not limited to this.

[0060] The learning unit 123 generates a first trained model by performing non-defective learning on the physical quantities acquired by the physical quantity acquisition unit 121. The first trained model is a trained model that receives as input the physical quantities generated in the power source when a new screw is tightened, and outputs whether the new screw tightening has been completed normally without any difference from the characteristics of normal screw tightening.

[0061] In the first embodiment, the learning unit 123 performs non-defective learning on n torque waveforms acquired by the physical quantity acquisition unit 121 to generate a first trained model. All of the n torque waveforms are normal torque waveforms when screw tightening is completed normally. For example, the first trained model in the first embodiment is a trained model that receives as input the torque waveform generated in the motor 30 when a new screw is tightened, and outputs whether the new screw tightening has been completed normally without differing from the characteristics of normal screw tightening.

[0062] In the first embodiment, the k-nearest neighbor algorithm (knn) is used as an example of a machine learning method for good product learning, but the present invention is not limited to this. Other examples of the machine learning method for good product learning include, but are not limited to, One Class SVM (Support Vector Machine) and Local Outlier Factor (LOF).

[0063] 6 and 7 are explanatory diagrams showing an example of a non-defective product learning method by the learning unit 123 of the first embodiment. As mentioned above, FIGS. 6 and 7 are explanatory diagrams of a non-defective product learning method using the k-nearest neighbor method. In the example shown in FIGS. 6 and 7, feature quantities FV1 to FVn of n torque waveforms, which are the torque waveforms of the normal system mentioned above, are represented in a high-dimensional space. The feature quantities of the torque waveform are represented, for example, as a vector of dimensions equal to the number of samples, in which the torque values ​​when the torque waveform is sampled at the above-mentioned sample intervals are arranged. For example, when the number of samples is m (m≧2), the feature quantity FV1 is expressed as FV1=(FV11, FV12, ..., FV1 m ), the feature FVn is FVn=(FVn1,FVn2,…,FVn m ) Similarly, the dimension of the high-dimensional space is the same as the number of samples.

[0064] In the good product learning method using the k-nearest neighbor method, the learning unit 123 calculates the distance from the closest feature to the k-th feature for each of n feature quantities, and sets the maximum distance among the calculated n distances as the threshold for determining a defective product. In the first embodiment, the case of k=3 will be described as an example, but the present invention is not limited to this. In addition, the distance may be Euclidean distance, but is not limited to this. For example, when x=(x1, x2, ..., x L ), y=(y1,y2,…,y L ), the Euclidean distance between x and y is expressed by Equation (1).

[0065]

number

[0066] The example shown in Fig. 6 shows the distance found for feature FV1. In the example shown in Fig. 6, feature amounts close to feature FV1 are FV4, FV2, FV5, FV3, ... in order of decreasing distance. For example, the learning unit 123 identifies feature FV5 as the kth (k=3) closest feature, and finds the distance k15 between feature FV1 and feature FV5.

[0067] 7 shows the distance calculated for the feature FV2. In the example shown in Fig. 7, the features closest to the feature FV2 are FV4, FV3, FV1, ... in descending order of distance. For example, the learning unit 123 identifies the feature FV1 as the kth (k=3) closest feature, and calculates the distance k21 between the feature FV2 and the feature FV1.

[0068] Although the explanation will be omitted below, the learning unit 123 similarly identifies the kth (k=3)th closest feature for each of the feature amounts FV3, FV4, ..., FVn, and calculates the distance from the identified feature amount. The learning unit 123 identifies the largest distance from the n distances calculated for the feature amounts FV1 to FVn, and sets this as the threshold for determining whether the feature amount is good or bad. The learning unit 123 sets the feature amounts FV1 to FVn, the value of k (k=3), and the calculated threshold for determining whether the feature amount is good or bad as the first trained model.

[0069] Next, we will explain the functions of the physical quantity acquisition unit 121, the model acquisition unit 125, and the judgment unit 127 when performing defect judgment in the defect judgment device 100. Note that the defect judgment is performed after the first trained model is generated by the above-mentioned non-defective product learning.

[0070] The physical quantity acquisition unit 121 acquires a target physical quantity, which is a physical quantity generated in the power source when a new screw is tightened. It is assumed that the size of the screw used for the new screw tightening is the same as the size of the screw used during good product learning. The physical quantity is at least one of the torque waveform and the rotational speed waveform of the power source. In the first embodiment, as in good product learning, a case will be described as an example in which the power source is the motor 30 and the physical quantity is the torque waveform of the motor 30, but the present invention is not limited to this.

[0071] The model acquisition unit 125 acquires a first trained model that has been obtained by performing non-defective learning on the physical quantities that occurred in the power source during a previous screw tightening operation that was successfully completed. In the first embodiment, the model acquisition unit 125 acquires the first trained model generated by the learning unit 123.

[0072] The determination unit 127 applies the target physical quantity acquired by the physical quantity acquisition unit 121 to the first trained model acquired by the model acquisition unit 125, and determines whether the new screw tightening has been completed normally without any discrepancy with the characteristics of normal screw tightening. The determination unit 127 outputs the defect determination result to the main control device 70. Note that the determination unit 127 may output the defect determination result to a display device via the input / output interface 107, or may transmit the defect determination result to the outside via the communication interface 109.

[0073] 8 is an explanatory diagram showing an example of a defect determination method using the first trained model by the determination unit 127 of the first embodiment. In the example shown in FIG. 8, the above-mentioned feature quantities FV1 to FVn are arranged in a high-dimensional space. In addition, the feature quantities NFV1 and NFV2 of the torque waveform acquired as the target physical quantity by the physical quantity acquisition unit 121 are also arranged in the high-dimensional space. Note that, like the feature quantities FV1 to FVn, the feature quantities NFV1 and NFV2 of the torque waveform acquired as the target physical quantity are expressed as vectors with dimensions equal to the number of samples, in which the torque values ​​when the torque waveform is sampled at the above-mentioned sample intervals are arranged.

[0074] In the defect determination method using the first trained model, the determination unit 127, as in the case of non-defective product learning, determines the distance from the feature acquired as the target physical quantity to the closest k (k = 3)th feature quantity. As in the case of non-defective product learning, the distance may be, but is not limited to, Euclidean distance. The determination unit 127 compares the determined distance with a defect determination threshold determined by the learning unit 123. If the determined distance is less than the defect determination threshold, the determination unit 127 determines that the new screw tightening has been completed normally without any difference from the characteristics of normal screw tightening. On the other hand, if the determined distance is equal to or greater than the defect determination threshold, the determination unit 127 determines that the new screw tightening corresponds to some kind of defect. Note that "the new screw tightening corresponds to some kind of defect" means a state in which the characteristics of the target physical quantity in the new screw tightening differ from the characteristics of the target physical quantity in a previous screw tightening that was completed normally, a state in which the characteristics of the new screw tightening differ from the characteristics of a previous screw tightening that was completed normally, or a state in which the new screw tightening differs from the characteristics of normal screw tightening.

[0075] The example shown in FIG. 8 shows the distance found for the feature NFV1. In the example shown in FIG. 8, the feature quantities closest to the feature NFV1 are FV1, FV4, FV6, FV5, ... in order of closest distance. For example, the determination unit 127 identifies the feature FV6 as the kth (k=3) feature quantity from the closest, and determines the distance kn16 from the feature NFV1 to the feature FV6. The determination unit 127 compares the found distance kn16 with a threshold value for determining a defect. Here, it is assumed that the found distance kn16 is less than the threshold value for determining a defect. Therefore, the determination unit 127 determines that the new screw tightening for which the feature NFV1 has been found has been completed normally without any difference from the characteristics of normal screw tightening.

[0076] The example shown in Fig. 8 also shows the distance found for the feature NFV2. In the example shown in Fig. 8, the feature FV6 is the kth (k = 3) feature closest to the feature NFV2. Therefore, the determination unit 127 finds the distance kn26 between the feature NFV2 and the feature FV6, and compares the found distance kn26 with the threshold for determining a defect. Here, it is assumed that the found distance kn26 is equal to or greater than the threshold for determining a defect. Therefore, the determination unit 127 determines that the new screw tightening for which the feature NFV2 has been found corresponds to some kind of defect.

[0077] Next, the processing flow of the screw fastening device 10 will be described.

[0078] FIG. 9 is a flowchart showing an example of a screw fastening process performed by the screw fastening device 10 of the first embodiment.

[0079] First, the main control device 70 moves the driver bit 20 to the screw tightening position so that the screw 5 attracted and held by the driver bit 20 is positioned directly above the screw hole 2 (step S101).

[0080] Next, main control device 70 lowers driver bit 20 and presses screw 5, which is sucked and held, into screw hole 2 (step S103).

[0081] Next, the main control device 70 operates the motor 30 to rotate the driver bit 20 (step S105).

[0082] In steps S103 to S105, main control device 70 may operate motor 30 to rotate driver bit 20 while lowering driver bit 20, and insert screw 5, which is held by suction, into screw hole 2.

[0083] Next, main control device 70 determines whether the torque generated by motor 30 has reached the target torque (step S107).

[0084] If the target torque has not been reached (No in step S107), the main control device 70 determines whether the maximum screw tightening time has been reached (step S109).

[0085] If the maximum screw tightening time has not been reached (No in step S109), the process returns to step S107. On the other hand, if the maximum screw tightening time has been reached (Yes in step S109), it is assumed that some kind of malfunction has occurred and screw tightening has not been completed normally, and the process ends.

[0086] On the other hand, if the target torque is reached in step S107 (Yes in step S107), the defect determination device 100 performs a learning / defect determination process (step S111).

[0087] Fig. 10 is a flowchart showing an example of a learning (good product learning) process by the defect determination device 100 of the first embodiment. Note that the process shown in Fig. 10 is performed in step S111 of Fig. 9 when n torque waveforms have been collected, in a case where the screw tightening process shown in Fig. 9 is performed for the purpose of collecting torque waveforms of normally completed screw tightening for good product learning. However, in a case where n torque waveforms of normally completed screw tightening have been prepared in advance, the process is performed independently of the process shown in Fig. 9.

[0088] First, the physical quantity acquisition unit 121 receives n normal torque waveforms and a parameter k (step S201).

[0089] Next, the learning unit 123 sets the value of i to 1 (step S203) and sets the value of j to 1 (step S205).

[0090] Next, the learning unit 123 calculates the Euclidean distance between the feature amount of the i-th torque waveform and the feature amount of the j-th torque waveform (step S207), and stores the calculated distance in memory as the j-th distance (step S209). The learning unit 123 repeats the processes of steps S205 to S209 until the process of step S209 when j is n is completed. However, the learning unit 123 skips the processes of steps S207 to S209 when i=j.

[0091] Next, the learning unit 123 sorts the n-1 distances stored in the memory in ascending order (step S211), and sets the kth smallest distance as the abnormality degree of the i-th torque waveform feature amount (step S213). The learning unit 123 repeats the processes of steps S203 to S213 until the process of step S213 when the value of i is n is completed.

[0092] Next, the learning unit 123 sets the maximum value of the set n abnormality degrees as the threshold value for determining a defect (step S215).

[0093] Fig. 11 is a flowchart showing an example of a defect determination process by the defect determination device 100 of the first embodiment. Note that the process shown in Fig. 11 is performed in step S111 of Fig. 9 when the screw tightening process shown in Fig. 9 is not performed for the purpose of collecting torque waveforms of normally completed screw tightening for good product learning.

[0094] First, the model acquisition unit 125 sets n normal torque waveforms, a parameter k, and a threshold value for defect determination as a first trained model (step S301).

[0095] Next, the physical quantity acquisition unit 121 inputs the waveform of the torque to be predicted as the target physical quantity (step S303).

[0096] Next, the determination unit 127 sets the value of i to 1 (step S305).

[0097] Next, the determination unit 127 calculates the Euclidean distance between the feature amount of the torque waveform to be predicted and the feature amount of the i-th torque waveform (step S307), and stores the calculated distance in memory as the i-th distance (step S309). The determination unit 127 repeats the processes of steps S305 to S309 until the process of step S309 when the value of i is n is completed.

[0098] Next, the determination unit 127 sorts the n distances stored in the memory in ascending order (step S311), and sets the kth smallest distance as the degree of abnormality of the feature amount of the torque waveform to be predicted (step S313).

[0099] Next, the determination unit 127 compares the set abnormality level with a threshold value for determining a defect, and outputs a result of the defect determination (step S315).

[0100] As described above, in the first embodiment, focusing on the fact that the torque waveform of an abnormal system tends to deviate from the torque waveform of a normal system, the defect determination device 100 generates a trained model that has learned the torque waveform of a normal system as a non-defective product. Using the trained model that has been generated, the defect determination device 100 determines whether the screw tightening by the screw tightening device 10 has been completed normally without any discrepancy with the characteristics of normal screw tightening. Therefore, according to the first embodiment, it is possible to determine whether the screw tightening by the screw tightening device 10 has been completed normally without any discrepancy with the characteristics of normal screw tightening, regardless of the type of defect.

[0101] For example, the defect determination device 100 of the first embodiment is not designed to detect a specific type of defect, and can therefore determine whether screw tightening has been completed normally for a variety of defects. For example, according to the first embodiment, in addition to defects that occur frequently, such as screw loosening, cam-out, and foreign object bite, it can also determine that defects that occur less frequently, such as foreign matter in the female thread, thread stripping, broken screws, incorrect threading, bolt neck spinning, missing threads, double tightening, galling (seizure), and bolt elongation, are defects.

[0102] Note that screw loosening is a defect in which a screw is inserted at an angle into the screw hole before being fastened into a fastened object. Cam-out is a defect in which the tip of the driver bit floats up and comes out of the screw hole. Foreign object entrapment is a defect in which a screw is fastened into a fastened object with debris or other particles caught in it. Foreign object in the female thread is a defect in which a screw is fastened into a fastened object with debris or other particles caught in the screw hole. Stripped screw is a defect in which the threads are stripped. Breakage screw is a defect in which the screw breaks. Wrong screw is a defect in which a screw with dimensions or shape different from the specifications is tightened. Freewheeling nose is a defect in which a screw is rotated without the screw being able to fit into the screw hole. Missing thread is a defect in which a screw is tightened into a screw hole without a thread present. Double tightening is a defect in which a screw that has already been tightened is tightened again. Galling (seizing) is a defect in which the screw seizes. Bolt elongation is a defect in which the screw stretches.

[0103] Furthermore, according to the first embodiment, even if an unintended defect occurs, it is possible to determine whether or not screw tightening has been completed normally.

[0104] (Second embodiment) In the second embodiment, an example of classifying defects in screw tightening where defects have occurred will be described. The following mainly describes the differences from the first embodiment, and components having the same functions as those in the first embodiment will be given the same names and symbols as those in the first embodiment, and their description will be omitted.

[0105] Fig. 12 is a block diagram showing an example of the functional configuration of the defect determination device 200 of the second embodiment. As shown in Fig. 12, the defect determination device 200 differs from the defect determination device 100 of the first embodiment in that it includes a classification unit 229, a physical quantity acquisition unit 221, a learning unit 223, and a model acquisition unit 225.

[0106] First, a description will be given of the functions of the physical quantity acquisition unit 221 and the learning unit 223 during learning performed in the defect determination device 200. In the defect determination device 200 of the second embodiment, a second learned model for defect classification is further learned.

[0107] The physical quantity acquisition unit 221 further acquires a physical quantity that occurred in the power source when the screw was tightened in which the abnormality occurred. The physical quantity acquisition unit 221 acquires the physical quantities of s (s≧2) screw tightenings in which the abnormality occurred. It is assumed that the same size screws are used in all s screw tightenings. The physical quantity is at least one of the torque waveform and the rotational speed waveform of the power source. In the second embodiment, the power source is also the motor 30 and the physical quantity is the torque waveform of the motor 30, but the present invention is not limited to this.

[0108] The learning unit 223 learns the physical quantities acquired by the physical quantity acquisition unit 221 and generates a second trained model. The second trained model is a trained model that takes as input the physical quantities generated in the power source when a new screw is tightened, and outputs a defect classification result for the new screw tightening.

[0109] In the second embodiment, the learning unit 223 learns the torque waveforms for s times acquired by the physical quantity acquisition unit 221 to generate a second trained model. The torque waveforms for s times are all abnormal torque waveforms when an abnormality occurs in screw tightening. For example, the second trained model is a trained model that receives as input the torque waveform generated in the motor 30 when a new screw is tightened, and outputs a classification result as to which defect the new screw tightening falls into.

[0110] In the second embodiment, a neural network is used as an example of the machine learning method for the second trained model, but the present invention is not limited to this. Other machine learning methods include, but are not limited to, a multiclass SVM (Support Vector Machine).

[0111] FIG. 13 is an explanatory diagram showing an example of a learning method of the second trained model by the learning unit 223 of the second embodiment. As described above, FIG. 13 is an explanatory diagram of a learning method using a neural network. In the example shown in FIG. 13, the neural network is a three-layer neural network consisting of input layer neurons (x1, x2, ..., xm), intermediate layer neurons (h1, h2, ..., hm), and output layer neurons (y1, y2, ..., ym). Note that m is the number of samples when sampling the torque waveform described in the first embodiment. However, the number of intermediate layer neurons and output layer neurons is not limited to m and can be set to any value. For example, the number of output layer neurons is preferably the number of defects of the type to be classified.

[0112] The input layer neurons are input with the feature quantity of the torque waveform of the abnormal system acquired by the physical quantity acquisition unit 221. As in the first embodiment, the feature quantity of the torque waveform is expressed as a vector of dimensions equal to the number of samples, in which the torque values ​​when the torque waveform is sampled at the aforementioned sample intervals are arranged. For example, when the number of samples is m (m≧2), the feature quantity AFV1 of the abnormal system is expressed as AFV1=(AFV11, AFV12, ..., AFV1 m ) The values ​​of the abnormal feature AFV1, AFV11, AFV12, ..., AFV1 m are input to the input layer neurons x1, x2, ..., xm, respectively.

[0113] Each of the output layer neurons y1, y2, ..., ym outputs the probability that the defect type assigned to it applies. For example, if the output layer neuron y1 is assigned to "screw loose," the output layer neuron y1 outputs the probability that the feature value of the torque waveform input to the input layer neuron corresponds to "screw loose." Also, for example, if the output layer neuron y2 is assigned to "cam-out," the output layer neuron y2 outputs the probability that the feature value of the torque waveform input to the input layer neuron corresponds to "cam-out."

[0114] The hidden layer neurons (h1, h2, …, hm) are weighted by the values ​​of the input layer neurons (x1, x2, …, xm). 1 11 , w 1 12 , w 1 1m ,...w 1 m1 , w 1 m2 , w 1 mm Similarly, the output layer neurons (y1, y2, ..., ym) multiply the values ​​of the hidden layer neurons (h1, h2, ..., hm) by the weights w 2 11 , w 2 12 , w 2 1m ,...w 2 m1 , w 2 m2 , w 2 mm is the value multiplied by

[0115] However, as mentioned above, the output layer neurons (y1, y2, ..., ym) output the probability of falling into the assigned defect type, and therefore the output value calculated with the initial weights will be far from the correct value. For this reason, the learning unit 223 adjusts (learns) the value of each weight so that the error between the output value and the correct value becomes small, thereby generating a network that can classify screw tightening defects, and sets this as the second trained model.

[0116] Next, a description will be given of the functions of the model acquisition unit 225 and the classification unit 229 when performing defect classification in the defect determination device 200. Note that the defect classification is performed after the second trained model is generated by the above-described learning.

[0117] The target physical quantity acquired by the physical quantity acquisition unit 221 and the non-defective product judgment performed by the judgment unit 127 are the same as those in the first embodiment.

[0118] The model acquisition unit 225 further acquires a second trained model that is trained using the physical quantity that occurred in the power source during a previous screw tightening operation in which an abnormality occurred as an input and the result of classifying the screw tightening as defective as an output. In the second embodiment, the model acquisition unit 225 acquires the second trained model generated by the learning unit 223.

[0119] When the determination unit 127 determines that the new screw tightening has not been completed normally, the classification unit 229 applies the target physical quantity to the second trained model acquired by the model acquisition unit 225 to classify the new screw tightening as defective. The classification unit 229 outputs the defect classification result to the main control device 70. Note that the classification unit 229 may output the defect classification result to a display device via the input / output interface 107, or may transmit the defect classification result to the outside via the communication interface 109.

[0120] For example, suppose that the determination unit 127 determines that a new screw tightening for which the feature NFV2 described in the first embodiment has been obtained corresponds to some kind of defect. In this case, the classification unit 229 inputs the feature NFV2 to the input layer neurons (x1, x2, ..., xm) and obtains an output of a classification result indicating whether the screw tightening is classified as a defect from the output layer neurons (y1, y2, ..., ym).

[0121] 14 is a flowchart showing an example of a defect classification process by the defect determination device 200 of the second embodiment. The process shown in FIG. 14 is performed following the defect determination in step S315 of the defect determination process shown in FIG.

[0122] If the judgment unit 127 judges that the new screw tightening was not completed normally and is defective (Yes in step S315-1), the classification unit 229 applies the target physical quantity to the second trained model to classify the new screw tightening as defective (step S315-2) and outputs the defect classification result (step S315-3).

[0123] On the other hand, if the judgment unit 127 judges that the new screw tightening has been completed normally (No in step S315-1), the judgment unit 127 outputs the defect judgment result indicating that the screw tightening has been completed normally (step S315-4).

[0124] In the first embodiment, if the maximum screw tightening time is reached in step S109 of the screw tightening process shown in Fig. 9 (Yes in step S109), the process is terminated. However, in the second embodiment, in order to classify the type of defect, the learning / defect determination process in step S111 is performed in this case as well.

[0125] Furthermore, although not described in the flowchart, when the screw tightening process shown in Fig. 9 is performed for the purpose of collecting abnormal torque waveforms for learning defect classification, once s torque waveforms have been collected, the process of generating the second trained model described in Fig. 13 is performed in step S111 of Fig. 9. However, if s abnormal torque waveforms have been prepared in advance, this process is performed independently of the process shown in Fig. 9.

[0126] As described above, in the second embodiment, when the screw fastening performed by the screw fastening device 10 is defective in some way, the type of the defect can be classified.

[0127] (Variation 1) In the second embodiment, a case where fault classification is performed using the second trained model on the premise of fault determination using the first trained model has been described, but this is not limited to this. For example, in an environment where torque waveforms of abnormal systems are prepared in advance, the fault classification device may be configured to perform fault classification using the second trained model alone.

[0128] (Variation 2) In the second embodiment, as described above, an example has been described in which defect classification is performed using the second trained model on the premise that defect determination is performed using the first trained model, but defect determination may also be performed using the second trained model. In this case, the neural network may be trained using training data that also includes a torque waveform of a normal system, and one neuron may be added to the output layer to output the probability that screw tightening has been completed successfully.

[0129] (Variation 3) In the first and second embodiments, the torque waveform has been used as an example of a physical quantity generated in the motor 30, but this is not limiting. The rotational speed waveform of the motor 30 can also be used in the same way as the torque waveform, since it reflects the characteristics of each screw tightening process. In the case of the rotational speed waveform of the motor 30, the rotational speed is maintained at a relatively high value until the screw 5 begins to be tightened into the screw hole 2. Then, when the screw reaches a pre-seated state, the rotational speed begins to decrease from a relatively high value. When the screw reaches a final tightening state, the rotational speed suddenly decreases and is maintained at a relatively low value. In the first and second embodiments, the physical quantity generated in the motor 30 may be a combination of a feature value of the torque waveform and a feature value of the rotational speed waveform.

[0130] (Third embodiment) In the third embodiment, an example will be described in which preprocessing is performed on the waveform of the torque to be determined in order to accommodate a variety of screw sizes. The following mainly describes the differences from the second embodiment, and components having the same functions as those in the second embodiment will be given the same names and symbols as those in the second embodiment, and their description will be omitted.

[0131] Fig. 15 is a block diagram showing an example of the functional configuration of a defect determination device 300 according to the third embodiment. As shown in Fig. 15, the defect determination device 300 differs from the defect determination device 200 according to the second embodiment in that it includes a preprocessing unit 331 and a physical quantity acquisition unit 321.

[0132] The physical quantity acquisition unit 321 acquires a target physical quantity, which is a physical quantity generated in the power source when a new screw is tightened. In the third embodiment, the size of the screw used for the new screw tightening is assumed to be different from the size of the screw used during learning. The screw size refers to at least one of the length and thickness of the screw. It is assumed that the size of the screw used during learning (the screw used for the previous screw tightening) is a screw of a predetermined size.

[0133] When the screw used in new screw tightening is not of a predetermined size, the pre-processing unit 331 normalizes at least a part of the target physical quantity so that the target physical quantity matches a physical quantity generated in the power source during previous screw tightening. The at least a part of the target physical quantity is, for example, a physical quantity from the alignment of the screw to the provisional seating of the screw.

[0134] FIG. 16 is an explanatory diagram showing an example of a preprocessing method for target physical quantities by the preprocessing unit 331 according to the third embodiment. The example shown in FIG. 16 shows an example of the waveform of torque generated in the motor 30 from the start to the end of screw tightening by the screw tightening device 10. Waveforms 401n and 401a show the waveform of torque when a screw of a predetermined length is tightened. Waveform 401n is the waveform of torque in a normal system, and waveform 401a is the waveform of torque in an abnormal system. In the example shown in FIG. 16, it is assumed that a first trained model is trained using waveform 401n and the like, and a second trained model is trained using waveform 401a and the like.

[0135] Furthermore, waveforms 403n and 403a represent torque waveforms when a screw shorter than a predetermined length is tightened. Waveform 403n is a normal torque waveform, and waveform 403a is an abnormal torque waveform. Furthermore, waveform 405n represents a torque waveform when a screw longer than a predetermined length is tightened. In the third embodiment, these waveforms are normalized so that even torque waveforms when a screw of a different length from the predetermined length, such as waveforms 403n, 403a, and 405n, are tightened as target physical quantities, can be used for fault determination and fault classification.

[0136] Here, the maximum torque value does not change depending on the length of the screw, but the time from the start of screw tightening (alignment) to temporary seating changes in proportion to the length of the screw. For this reason, as shown in Figure 16, even if the screw length is different, the shape of the waveform from temporary seating to the end of screw tightening (torque maintenance) does not change much, but the shape of the waveform from the start of screw tightening to temporary seating contracts or expands in the time direction depending on the screw size.

[0137] For this reason, in the third embodiment, for torque waveforms during tightening of screws shorter than a predetermined length, such as waveforms 403n and 403a, the waveforms from the start of tightening to provisional seating are normalized so that they become longer in the time direction. Specifically, the first sample number, which is the number of samples from the start of tightening to provisional seating for waveforms 403n and 403a, is normalized to be equal to the specified sample number, which is the number of samples from the start of tightening to provisional seating for waveforms 401n and 401a.

[0138] Similarly, for the torque waveform during tightening of a screw longer than the predetermined length, such as waveform 405n, the waveform from the start of tightening to the provisional seating is normalized so that it becomes shorter in the time direction. Specifically, the second sample number, which is the number of samples from the start of tightening to the provisional seating of waveform 405n, is normalized to be the specified number of samples from the start of tightening to the provisional seating of waveforms 401n and 401a.

[0139] This makes it possible to make a defect judgment and classify a defect even when a torque waveform of a screw having a length different from a predetermined length is input as the target physical quantity.

[0140] Regarding the waveform from the provisional seating to the end of screw tightening, if the difference in the number of samples is on the order of an error, normalization in the time direction is not required.

[0141] FIG. 17 is a flowchart showing an example of advance preparation processing for performing preprocessing by the defect determination device 300 of the third embodiment.

[0142] First, in order to deal with a screw of a new length that is different from a screw of a predetermined length, the physical quantity acquisition unit 321 inputs t normal torque waveforms when the screw is tightened with the screw of the new length (step S401). Next, the pre-processing unit 331 detects the time point of temporary seating for each of the t input normal torque waveforms (step S403). Next, the pre-processing unit 331 calculates an average temporary seating time, which is the average of the temporary seating times from the start of screw tightening (alignment) to temporary seating, for each of the t input normal torque waveforms (step S405).

[0143] FIG. 18 is a flowchart showing an example of pre-processing by the defect determination device 300 of the third embodiment.

[0144] First, the pre-processing unit 331 sets the average temporary seating time calculated in the advance preparation process and the specified number of samples from the start of screw tightening (alignment) to temporary seating when tightening a screw of a predetermined length (step S501).

[0145] Next, the physical quantity acquisition unit 321 inputs the torque waveform during new screw tightening using a screw of a new length as the torque waveform to be predicted (step S503).

[0146] Next, the pre-processing unit 331 divides the input torque waveform to be predicted into a waveform from the start of screw tightening (alignment) to before temporary seating, and a waveform from after temporary seating to the end of screw tightening (torque maintenance) using the average temporary seating time (step S505).

[0147] Next, if the number of waveform samples from the start of screw tightening (alignment) to before temporary seating is less than the specified number of samples (Yes in step S507), the pre-processing unit 331 interpolates the intervals between samples so that the number of samples is equal to the specified number of samples (step S509).

[0148] On the other hand, if the number of waveform samples from the start of screw tightening (alignment) to before temporary seating is greater than the specified number of samples (No in step S507), the pre-processing unit 331 reduces the number of sample points so that it becomes equal to the specified number of samples (step S511).

[0149] Next, if the number of waveform samples from after temporary seating to the end of screw tightening (torque maintenance) is less than the specified number of samples (Yes in step S513), the pre-processing unit 331 interpolates the intervals between samples so that the number of samples is equal to the specified number of samples (step S515).

[0150] On the other hand, if the number of waveform samples from after temporary seating to the end of screw tightening (torque maintenance) is greater than the specified number of samples (No in step S513), the pre-processing unit 331 reduces the number of sample points so that it is equal to the specified number of samples (step S517).

[0151] As described above, in the third embodiment, since the change in the torque waveform when the screw length is different differs before and after temporary seating, normalization is also performed separately before and after temporary seating, with different degrees of normalization being performed for each. Therefore, according to the third embodiment, it is possible to perform defect judgment and defect classification corresponding to a variety of screw sizes.

[0152] (Variation 4) In the above third embodiment, the torque waveform was used as an example of the physical quantity generated in the motor 30, but when the length of the screw is different, a similar tendency occurs in the waveform of the rotation speed of the motor 30. For this reason, the waveform of the rotation speed of the motor 30 can also be used in the same way as the torque waveform.

[0153] (Variation 5) In the third embodiment, pre-processing for different screw lengths was described, but the method described in the third embodiment can also be applied to different screw thicknesses. However, when the screw thickness is different, the time required for the entire screw tightening process, not just the temporary seating process, changes. Furthermore, when the screw thickness is different, the torque required for the entire screw tightening process also changes. For this reason, when the screw thickness is different, normalization in the time direction and normalization in the torque direction are taken into consideration.

[0154] (program) The programs executed by the defect judgment devices 100, 200, and 300 of each of the above embodiments and each of the above variations are provided as files stored in an installable or executable format on a computer-readable storage medium such as a CD-ROM, CD-R, memory card, DVD, or flexible disk (FD).

[0155] The programs executed by the defect determination devices 100, 200, and 300 of the above embodiments and modifications may be stored on a computer connected to a network such as the Internet and provided by being downloaded via the network. The programs executed by the defect determination devices 100, 200, and 300 of the above embodiments and modifications may be provided or distributed via a network such as the Internet. The programs executed by the defect determination devices 100, 200, and 300 of the above embodiments and modifications may be provided by being pre-installed in a ROM or the like.

[0156] The programs executed by the defect determination devices 100, 200, and 300 of the above-described embodiments and modifications have a modular configuration for implementing the above-described units on a computer. As for actual hardware, for example, the CPU reads the learning program from the HDD onto the RAM and executes it, thereby implementing the above-described units on a computer.

[0157] As described above, according to the above-described embodiments and modifications, it is possible to determine whether or not screw tightening has been completed normally, regardless of the type of defect.

[0158] The above-described embodiments and modifications merely illustrate examples of specific embodiments of the present disclosure, and the technical scope of the present disclosure should not be construed as being limited by these. Therefore, the present disclosure can be implemented in various forms without departing from the spirit or main features thereof. For example, the above-described embodiments and modifications may be appropriately combined in their respective constituent units. Furthermore, for example, some components may be deleted from all components in the above-described embodiments and modifications.

[0159] In the above description, the notation "... part" used for each component may be replaced with other notations such as "... assembly," "... circuit," "... device," "... unit," or "... module."

[0160] The present disclosure includes the following aspects.

[0161] (1) a physical quantity acquisition unit that acquires a target physical quantity, which is a physical quantity generated in the power source when a new screw is tightened; a model acquisition unit that acquires a first trained model that has undergone good product learning of a physical quantity that occurred in the power source during a past screw tightening that was successfully completed; a determination unit that applies the target physical quantity to the first trained model and determines whether the new screw tightening has been completed successfully; A defect determination device comprising:

[0162] (2) The model acquisition unit further acquires a second trained model trained using a physical quantity that occurred in the power source during a previous screw tightening operation in which an abnormality occurred as an input and a fault classification result of the screw tightening operation as an output; The defect determination device described in (1) further includes a classification unit that, when it is determined that the new screw tightening has not been completed normally, applies the target physical quantity to the second trained model to classify the new screw tightening as defective.

[0163] (3) The screws used in the previous screw tightening were of a predetermined size, The defect determination device according to (1), further comprising a pre-processing unit that, when a screw used in the new screw tightening is not of the predetermined size, normalizes at least a part of the target physical quantity so that the target physical quantity matches a physical quantity generated in the power source during the previous screw tightening.

[0164] (4) The defect determination device according to (3), wherein at least a portion of the target physical quantities is a physical quantity from screw alignment to temporary seating among the target physical quantities.

[0165] (5) The defect determination device according to any one of (1) to (4), wherein the physical quantity is at least one of a waveform of torque and a waveform of rotation speed of the power source.

[0166] (6) a physical quantity acquisition unit that acquires a target physical quantity, which is a physical quantity generated in the power source when a new screw is tightened; a model acquisition unit that acquires a trained model that uses as input a physical quantity that occurred in the power source during a previous screw tightening operation in which an abnormality occurred, and as output a fault classification result of the screw tightening operation; a classification unit that applies the target physical quantity to the trained model to classify the new screw tightening defect; A defect classification device comprising:

[0167] (7) a physical quantity acquisition unit that acquires a physical quantity generated in the power source when the screw tightening is successfully completed; a learning unit that performs non-defective learning of the physical quantity and generates a trained model that inputs a physical quantity generated in the power source during new screw tightening and outputs whether the new screw tightening has been completed normally; A model generation device comprising:

[0168] (8) a physical quantity acquisition unit that acquires a physical quantity that occurs in the power source when an abnormality occurs in screw tightening; a learning unit that learns learning data including the physical quantity, and generates a trained model that inputs the physical quantity generated in the power source during new screw tightening and outputs a defect classification result for the new screw tightening; A model generation device comprising:

[0169] (9) a physical quantity acquisition step in which the physical quantity acquisition unit acquires a target physical quantity, which is a physical quantity generated in the power source during new screw tightening; a model acquisition step in which a model acquisition unit acquires a first trained model obtained by non-defective learning of a physical quantity that occurred in the power source during a past screw tightening that was successfully completed; a determination step in which a determination unit applies the target physical quantity to the first trained model and determines whether the new screw tightening has been completed normally; A method for determining defects, including:

[0170] (10) a physical quantity acquisition step in which the physical quantity acquisition unit acquires a target physical quantity, which is a physical quantity generated in the power source during new screw tightening; a model acquisition step in which a model acquisition unit acquires a trained model that is trained using a physical quantity that occurred in the power source during a past screw tightening in which an abnormality occurred as an input and a fault classification result of the screw tightening as an output; a classification step in which a classification unit applies the target physical quantity to the trained model to classify the new screw tightening defect; A defect classification method including:

[0171] (11) a physical quantity acquisition step in which the physical quantity acquisition unit acquires a physical quantity generated in the power source when the screw tightening is successfully completed; a learning step in which a learning unit performs non-defective learning of the physical quantity and generates a trained model in which a physical quantity generated in the power source during new screw tightening is input and an output is output as to whether the new screw tightening has been completed normally; A model generation method including:

[0172] (12) a physical quantity acquisition step in which the physical quantity acquisition unit acquires a physical quantity that occurred in the power source during screw tightening in which an abnormality occurred; a learning step in which a learning unit learns learning data including the physical quantity and generates a trained model that inputs a physical quantity generated in the power source during new screw tightening and outputs a defect classification result for the new screw tightening; A model generation method including:

[0173] (13) A trained model generated by the model generation device according to (7) or (8), or the model generation method according to (11) or (12).

[0174] (14) A program for causing a computer to execute the method described in any one of (9) to (12). [Explanation of symbols]

[0175] 1 Work 2 screw holes 5 screws 10 Screw tightening device 20 screwdriver bits 30 motors 40 stages 50 Motor Drive Unit 60 Stage drive unit 70 Main control unit 100, 200, 300 defective judgment device 101 processors 103 memory 105 Auxiliary storage device 107 Input / Output Interface 109 Communication Interface 111 Various buses 121, 221, 321 Physical quantity acquisition part 123, 223 Learning Department 125, 225 Model Acquisition Department 127 Judgment section 229 Classification Department 331 Pretreatment section

Claims

1. a physical quantity acquisition unit that acquires a target physical quantity that is a physical quantity generated in the power source when a new screw is tightened; a model acquisition unit that acquires a first trained model that has undergone good product learning of a physical quantity that occurred in the power source during a past screw tightening that was successfully completed; a determination unit that applies the target physical quantity to the first trained model and determines whether the new screw tightening has been completed successfully; A defect determination device comprising:

2. The model acquisition unit further acquires a second trained model that is trained using a physical quantity that occurred in the power source during a past screw tightening operation in which an abnormality occurred as an input and a fault classification result of the screw tightening operation as an output, 2. The defect determination device according to claim 1, further comprising a classification unit that, when it is determined that the new screw tightening has not been completed normally, applies the target physical quantity to the second trained model to classify the new screw tightening as defective.

3. The screw used in the previous screw tightening is a screw of a predetermined size, 2. The defect determination device according to claim 1, further comprising: a pre-processing unit that, when a screw used in the new screw tightening is not of the predetermined size, normalizes at least a part of the target physical quantity so that the target physical quantity matches a physical quantity generated in the power source during the previous screw tightening.

4. The defect determination device according to claim 3 , wherein at least a portion of the target physical quantities is a physical quantity from a screw alignment to a screw provisional seating among the target physical quantities.

5. 5. The defect determination device according to claim 1, wherein the physical quantity is at least one of a waveform of torque and a waveform of rotation speed of the power source.

6. a physical quantity acquisition unit that acquires a target physical quantity that is a physical quantity generated in the power source when a new screw is tightened; a model acquisition unit that acquires a trained model that uses as input a physical quantity that occurred in the power source during a previous screw tightening operation in which an abnormality occurred, and as output a fault classification result of the screw tightening operation; a classification unit that applies the target physical quantity to the trained model to classify the new screw tightening defect; A defect classification device comprising:

7. a physical quantity acquisition unit that acquires a physical quantity generated in the power source when the screw tightening is successfully completed; a learning unit that performs non-defective learning of the physical quantity and generates a trained model that inputs a physical quantity generated in the power source during new screw tightening and outputs whether the new screw tightening has been completed normally; A model generation device comprising:

8. a physical quantity acquisition unit that acquires a physical quantity that occurs in the power source when an abnormality occurs in screw tightening; a learning unit that learns learning data including the physical quantity, and generates a trained model that inputs the physical quantity generated in the power source during new screw tightening and outputs a defect classification result for the new screw tightening; A model generation device comprising:

9. a physical quantity acquisition step in which the physical quantity acquisition unit acquires a target physical quantity, which is a physical quantity generated in the power source during new screw tightening; a model acquisition step in which the model acquisition unit acquires a first trained model obtained by non-defective learning of a physical quantity that occurred in the power source during a past screw tightening that was successfully completed; a determination step in which a determination unit applies the target physical quantity to the first trained model and determines whether the new screw tightening has been completed normally; A method for determining defects, including:

10. a physical quantity acquisition step in which the physical quantity acquisition unit acquires a target physical quantity, which is a physical quantity generated in the power source during new screw tightening; a model acquisition step in which a model acquisition unit acquires a trained model that is trained using a physical quantity that occurred in the power source during a past screw tightening in which an abnormality occurred as an input and a fault classification result of the screw tightening as an output; a classification step in which a classification unit applies the target physical quantity to the trained model to classify the new screw tightening defect; A defect classification method including:

11. a physical quantity acquisition step in which the physical quantity acquisition unit acquires a physical quantity generated in the power source when the screw tightening is successfully completed; a learning step in which a learning unit performs non-defective learning of the physical quantity and generates a trained model in which a physical quantity generated in the power source during new screw tightening is input and an output is output as to whether the new screw tightening has been completed normally; A model generation method including:

12. a physical quantity acquisition step in which a physical quantity acquisition unit acquires a physical quantity that occurred in the power source during screw tightening in which an abnormality occurred; a learning step in which a learning unit learns learning data including the physical quantity, and generates a trained model in which a physical quantity generated in the power source during new screw tightening is input and a fault classification result of the new screw tightening is output; A model generation method including:

13. A trained model generated by the model generation device according to claim 7 or 8, or the model generation method according to claim 11 or 12.

14. A program for causing a computer to execute the method according to any one of claims 9 to 12.

Images