Control device for injection molding machine, management device for injection molding machine, display device, injection molding machine, and machine learning device.

Abstract

It easily detects abnormalities. [Solution] A control device for an injection molding machine according to one embodiment acquires first time-series data showing the results detected by a detection device provided in the injection molding machine in a time series for each shot in which a molded product is formed by the injection molding machine, and when the first time-series data is input to a machine learning model that has been machine-learned to estimate the time series of the results detected by the detection device for each shot in the injection molding machine, the control unit receives output data inferred from the machine learning model as the time-series data of the results detected by the detection device in the time period corresponding to the first time-series data, and performs control to detect an anomaly based on the difference between the first time-series data and the output data.

Description

[Technical Field]

[0001] The present invention relates to a control device for an injection molding machine, a management device for an injection molding machine, a display device, an injection molding machine, and a machine learning device. [Background technology]

[0002] Conventional injection molding machines are equipped with a monitoring function that performs monitoring based on detection results obtained from sensors built into or connected to the injection molding machine. One example of such a monitoring function is a technology in which the user sets a threshold for statistical values ​​or representative values ​​of the detection results, and determines whether the statistical values ​​or representative values ​​exceed the threshold. However, in this technology, it is difficult for the user to intuitively understand what the statistical values ​​or representative values ​​are, making it difficult for the user to appropriately set the threshold for the statistical values ​​or representative values ​​according to the molding status.

[0003] In recent years, technologies using neural networks have been proposed for monitoring injection molding machines. For example, Patent Document 1 proposes a technology that acquires multiple time-series data based on detection signals from multiple sensors, processes the time-series data, and then uses it as training data to generate an estimation model through deep learning. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-75385 [Overview of the project] [Problems that the invention aims to solve]

[0005] The technology described in Patent Document 1 enables inference by assigning "good" and "bad" labels to the training data. However, in the technology described in Patent Document 1, it is necessary to train the model using various time-series data that represent molding defects in order to detect defects, which results in a heavy workload.

[0006] One aspect of the present invention provides a technology that enables easy detection of anomalies using machine learning models. [Means for solving the problem]

[0007] A control device for an injection molding machine according to one aspect of the present invention includes a control unit which, for each shot in which a molded product is formed by the injection molding machine, acquires first time-series data showing the results detected by a detection device provided in the injection molding machine in a time-series manner, and when the first time-series data is input to a machine learning model that has been machine-learned to estimate the time-series data of the results detected by the detection device for each shot in the injection molding machine, the control unit which receives output data inferred from the machine learning model as the time-series data of the results detected by the detection device in the time period corresponding to the first time-series data, and performs control to detect an anomaly based on the difference between the first time-series data and the output data. [Effects of the Invention]

[0008] According to one aspect of the present invention, abnormalities can be easily detected. [Brief explanation of the drawing]

[0009] [Figure 1] This figure shows the state of an injection molding machine upon completion of mold opening according to one embodiment. [Figure 2] This figure shows the state of an injection molding machine during mold clamping according to one embodiment. [Figure 3] This is a conceptual diagram illustrating the cooperation between the machine learning apparatus and the control device of the injection molding machine according to the first embodiment, specifically regarding the trained model. [Figure 4]It is a diagram showing an example of the functional configuration of a machine learning device according to the first embodiment. [Figure 5] It is a diagram showing an example of waveform data of pressure actual values which are detection results by a load detector provided in a test injection molding machine or an injection molding machine according to the first embodiment. [Figure 6] It is a diagram exemplifying pressure actual values which are detection results by a load detector provided in a test injection molding machine or an injection molding machine according to the first embodiment in a table structure. [Figure 7] It is a diagram explaining the concept of machine learning by a learning unit of a machine learning device according to the first embodiment. [Figure 8] It is a diagram showing an example of the functional configuration of a control device of an injection molding machine according to the first embodiment. [Figure 9] It is a diagram explaining the concept of processing executed by an abnormality detection unit and a display control unit according to the first embodiment. [Figure 10] It is an explanatory diagram showing a residual calculated by an abnormality detection unit according to the first embodiment. [Figure 11] It is a graph showing pressure actual values, pressure estimated values, and residuals at shot number "1004" by an injection molding machine according to the first embodiment. [Figure 12] It is a diagram showing an example of a screen displayed on a display device by a display control unit according to the first embodiment. [Figure 13] It is a diagram exemplifying a log information screen output by a display control unit according to the first embodiment. [Figure 14] It is a diagram exemplifying another aspect of a log information screen output by a display control unit according to the first embodiment. [Figure 15] It is a diagram exemplifying the configurations of a machine learning device, a group management device, and an injection molding machine according to the second embodiment.

Modes for Carrying Out the Invention

[0010] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In addition, the embodiments described below are illustrative and do not limit the invention, and not all features and combinations thereof described in the embodiments are necessarily essential to the invention. In each drawing, the same or corresponding components are denoted by the same or corresponding reference numerals, and the description may be omitted.

[0011] FIG. 1 is a view showing a state when the mold of the injection molding machine according to the first embodiment is fully opened. FIG. 2 is a view showing a state when the mold of the injection molding machine according to the first embodiment is clamped. In this specification, the X-axis direction, the Y-axis direction, and the Z-axis direction are perpendicular to each other. The X-axis direction and the Y-axis direction represent the horizontal direction, and the Z-axis direction represents the vertical direction. When the mold clamping device 100 is a horizontal mold, the X-axis direction is the mold opening / closing direction, and the Y-axis direction is the width direction of the injection molding machine 10. The negative side in the Y-axis direction is called the operation side, and the positive side in the Y-axis direction is called the non-operation side.

[0012] As shown in FIGS. 1 to 2, the injection molding machine 10 includes a mold clamping device 100 that opens and closes the mold device 800, an ejector device 200 that ejects the molded product formed by the mold device 800, an injection device 300 that injects a molding material into the mold device 800, a moving device 400 that moves the injection device 300 forward and backward with respect to the mold device 800, a control device 700 that controls each component of the injection molding machine 10, and a frame 900 that supports each component of the injection molding machine 10. The frame 900 includes a mold clamping device frame 910 that supports the mold clamping device 100 and an injection device frame 920 that supports the injection device 300. The mold clamping device frame 910 and the injection device frame 920 are respectively installed on the floor 2 via leveling adjusters 930. The control device 700 is disposed in the internal space of the injection device frame 920. Hereinafter, each component of the injection molding machine 10 will be described.

[0013] (Mold clamping device) In the description of the mold clamping device 100, the moving direction of the movable platen 120 when the mold is closed (for example, the positive X-axis direction) is described as the front, and the moving direction of the movable platen 120 when the mold is opened (for example, the negative X-axis direction) is described as the rear.

[0014] The mold clamping device 100 performs mold closing, pressure increasing, mold clamping, pressure decreasing, and mold opening of the mold apparatus 800. The mold apparatus 800 includes a fixed mold 810 and a movable mold 820. The mold clamping device 100 is, for example, a horizontal type, and the mold opening and closing direction is horizontal. The mold clamping device 100 has a fixed platen 110 to which the fixed mold 810 is attached, a movable platen 120 to which the movable mold 820 is attached, and a moving mechanism 102 that moves the movable platen 120 relative to the fixed platen 110 in the mold opening and closing direction.

[0015] The fixed platen 110 is fixed to the clamping device frame 910. The fixed mold 810 is attached to the surface of the fixed platen 110 facing the movable platen 120.

[0016] The movable platen 120 is positioned to move freely in the mold opening and closing direction relative to the mold clamping device frame 910. Guides 101 are laid on the mold clamping device frame 910 to guide the movable platen 120. A movable mold 820 is attached to the surface of the movable platen 120 facing the fixed platen 110.

[0017] The moving mechanism 102 performs mold closing, pressure increasing, mold clamping, depressurization, and mold opening of the mold device 800 by moving the movable platen 120 forward and backward relative to the fixed platen 110. The moving mechanism 102 includes a toggle support 130 positioned at a distance from the fixed platen 110, a tie bar 140 connecting the fixed platen 110 and the toggle support 130, a toggle mechanism 150 that moves the movable platen 120 in the mold opening and closing direction relative to the toggle support 130, a mold clamping motor 160 that operates the toggle mechanism 150, a motion conversion mechanism 170 that converts the rotational motion of the mold clamping motor 160 into linear motion, and a mold thickness adjustment mechanism 180 that adjusts the distance between the fixed platen 110 and the toggle support 130.

[0018] The toggle support 130 is positioned at a distance from the fixed platen 110 and is mounted on the mold clamping device frame 910 so as to be movable in the mold opening and closing direction. The toggle support 130 may also be positioned so as to be movable along a guide laid on the mold clamping device frame 910. The guide for the toggle support 130 may be the same as the guide 101 for the movable platen 120.

[0019] In this embodiment, the fixed platen 110 is fixed to the clamping device frame 910, and the toggle support 130 is arranged to be movable relative to the clamping device frame 910 in the mold opening and closing direction. However, the toggle support 130 may be fixed to the clamping device frame 910, and the fixed platen 110 may be arranged to be movable relative to the clamping device frame 910 in the mold opening and closing direction.

[0020] The tie bars 140 connect the fixed platen 110 and the toggle support 130 at a distance L in the mold opening and closing direction. Multiple tie bars 140 (for example, four) may be used. Multiple tie bars 140 are arranged parallel to the mold opening and closing direction and stretch in accordance with the clamping force. At least one tie bar 140 may be provided with a tie bar strain detector 141 that detects the strain of the tie bar 140. The tie bar strain detector 141 sends a signal indicating its detection result to the control device 700. The detection result of the tie bar strain detector 141 is used for detecting the clamping force, etc.

[0021] In this embodiment, a tie bar strain detector 141 is used as a clamping force detector to detect the clamping force, but the present invention is not limited to this. The clamping force detector is not limited to strain gauge type, but may be piezoelectric, capacitive, hydraulic, electromagnetic, etc., and its mounting position is not limited to the tie bar 140.

[0022] The toggle mechanism 150 is positioned between the movable platen 120 and the toggle support 130, and moves the movable platen 120 in the mold opening and closing direction relative to the toggle support 130. The toggle mechanism 150 has a crosshead 151 that moves in the mold opening and closing direction, and a pair of link groups that bend and extend as the crosshead 151 moves. Each of the link groups has a first link 152 and a second link 153 that are bendable and extendable connected by a pin or the like. The first link 152 is pivotably attached to the movable platen 120 by a pin or the like. The second link 153 is pivotably attached to the toggle support 130 by a pin or the like. The second link 153 is attached to the crosshead 151 via a third link 154. When the crosshead 151 moves forward and backward relative to the toggle support 130, the first link 152 and the second link 153 bend and extend, and the movable platen 120 moves forward and backward relative to the toggle support 130.

[0023] Note that the configuration of the toggle mechanism 150 is not limited to the configuration shown in Figures 1 and 2. For example, in Figures 1 and 2, each link group has five nodes, but it may also have four, and one end of the third link 154 may be connected to the node between the first link 152 and the second link 153.

[0024] The clamping motor 160 is attached to the toggle support 130 and operates the toggle mechanism 150. The clamping motor 160 moves the crosshead 151 forward and backward relative to the toggle support 130, thereby bending and extending the first link 152 and the second link 153, and moving the movable platen 120 forward and backward relative to the toggle support 130. The clamping motor 160 is directly connected to the motion conversion mechanism 170, but it may also be connected to the motion conversion mechanism 170 via a belt, pulley, or the like.

[0025] The motion conversion mechanism 170 converts the rotational motion of the clamping motor 160 into the linear motion of the crosshead 151. The motion conversion mechanism 170 includes a screw shaft and a screw nut that screws onto the screw shaft. A ball or roller may be interposed between the screw shaft and the screw nut.

[0026] The mold clamping device 100 performs processes such as mold closing, pressure boosting, mold clamping, depressurization, and mold opening under the control of the control device 700.

[0027] In the mold closing process, the clamping motor 160 is driven to advance the crosshead 151 to the mold closing completion position at a set movement speed, thereby advancing the movable platen 120 and bringing the movable mold 820 into contact with the fixed mold 810. The position and movement speed of the crosshead 151 are detected using, for example, a clamping motor encoder 161. The clamping motor encoder 161 detects the rotation of the clamping motor 160 and sends a signal indicating the detection result to the control device 700.

[0028] Furthermore, the crosshead position detector for detecting the position of the crosshead 151 and the crosshead speed detector for detecting the movement speed of the crosshead 151 are not limited to the clamping motor encoder 161, and general-purpose devices can be used. Similarly, the movable platen position detector for detecting the position of the movable platen 120 and the movable platen speed detector for detecting the movement speed of the movable platen 120 are not limited to the clamping motor encoder 161, and general-purpose devices can be used.

[0029] In the boosting process, the clamping motor 160 is further driven to advance the crosshead 151 from the closed position to the clamping position, thereby generating clamping force.

[0030] In the clamping process, the clamping motor 160 is driven to maintain the position of the crosshead 151 in the clamping position. In the clamping process, the clamping force generated in the pressurization process is maintained. In the clamping process, a cavity space 801 (see Figure 2) is formed between the movable mold 820 and the fixed mold 810, and the injection unit 300 fills the cavity space 801 with liquid molding material. A molded product is obtained when the filled molding material solidifies.

[0031] The number of cavity spaces 801 may be one or more. In the latter case, multiple molded products can be obtained simultaneously. An insert material may be placed in part of the cavity space 801, and the molding material may be filled in the other part of the cavity space 801. A molded product in which the insert material and the molding material are integrated is obtained.

[0032] In the depressurization process, the clamping motor 160 is driven to retract the crosshead 151 from the clamping position to the mold opening start position, thereby retracting the movable platen 120 and reducing the clamping force. The mold opening start position and the mold closing completion position may be the same position.

[0033] In the mold opening process, the clamping motor 160 is driven to retract the crosshead 151 from the mold opening start position to the mold opening completion position at a set movement speed, thereby retracting the movable platen 120 and separating the movable mold 820 from the fixed mold 810. Subsequently, the ejector device 200 ejects the molded product from the movable mold 820.

[0034] The setting conditions for the mold closing process, the pressure boosting process, and the mold clamping process are set together as a series of setting conditions. For example, the movement speed and position of the crosshead 151 (including the mold closing start position, movement speed switching position, mold closing completion position, and mold clamping position), and the mold clamping force in the mold closing process and the pressure boosting process are set together as a series of setting conditions. The mold closing start position, movement speed switching position, mold closing completion position, and mold clamping position are arranged in this order from rear to front and represent the start and end points of the section in which the movement speed is set. The movement speed is set for each section. There may be one or more movement speed switching positions. There may be no movement speed switching positions. The mold clamping position and the mold clamping force may be set individually or individually.

[0035] The setting conditions for the depressurization process and the mold opening process are set similarly. For example, the movement speed and position of the crosshead 151 in the depressurization process and the mold opening process (mold opening start position, movement speed switching position, and mold opening completion position) are set together as a series of setting conditions. The mold opening start position, movement speed switching position, and mold opening completion position are arranged in this order from front to back and represent the start and end points of the sections in which the movement speed is set. The movement speed is set for each section. There may be one or more movement speed switching positions. There may be no movement speed switching positions. The mold opening start position and the mold closing completion position may be the same position. Also, the mold opening completion position and the mold closing start position may be the same position.

[0036] Alternatively, the movement speed and position of the movable platen 120 may be set instead of the movement speed and position of the crosshead 151. Furthermore, the clamping force may be set instead of the position of the crosshead (e.g., the clamping position) or the position of the movable platen.

[0037] Incidentally, the toggle mechanism 150 amplifies the driving force of the clamping motor 160 and transmits it to the movable platen 120. This amplification ratio is also called the toggle ratio. The toggle ratio changes depending on the angle θ between the first link 152 and the second link 153 (hereinafter also referred to as the "link angle θ"). The link angle θ can be determined from the position of the crosshead 151. The toggle ratio is maximized when the link angle θ is 180°.

[0038] If the thickness of the mold device 800 changes due to replacement of the mold device 800 or a change in the temperature of the mold device 800, the mold thickness is adjusted so that a predetermined clamping force is obtained during mold clamping. In mold thickness adjustment, for example, the distance L between the fixed platen 110 and the toggle support 130 is adjusted so that the link angle θ of the toggle mechanism 150 becomes a predetermined angle at the time of mold touch when the movable mold 820 touches the fixed mold 810.

[0039] The mold clamping device 100 has a mold thickness adjustment mechanism 180. The mold thickness adjustment mechanism 180 adjusts the mold thickness by adjusting the distance L between the fixed platen 110 and the toggle support 130. The timing of the mold thickness adjustment is, for example, between the end of one molding cycle and the start of the next molding cycle. The mold thickness adjustment mechanism 180 includes, for example, a screw shaft 181 formed at the rear end of the tie bar 140, a screw nut 182 that is rotatably and immovably held by the toggle support 130, and a mold thickness adjustment motor 183 that rotates the screw nut 182 that is screwed onto the screw shaft 181.

[0040] A screw shaft 181 and screw nut 182 are provided for each tie bar 140. The rotational driving force of the mold thickness adjustment motor 183 may be transmitted to multiple screw nuts 182 via a rotational driving force transmission unit 185. Multiple screw nuts 182 can be rotated synchronously. It is also possible to rotate multiple screw nuts 182 individually by changing the transmission path of the rotational driving force transmission unit 185.

[0041] The rotational drive force transmission unit 185 is composed of, for example, gears. In this case, driven gears are formed on the outer circumference of each screw nut 182, a drive gear is attached to the output shaft of the mold thickness adjustment motor 183, and an intermediate gear that meshes with the multiple driven gears and the drive gear is rotatably held in the center of the toggle support 130. Note that the rotational drive force transmission unit 185 may be composed of a belt or pulley instead of gears.

[0042] The operation of the mold thickness adjustment mechanism 180 is controlled by the control device 700. The control device 700 drives the mold thickness adjustment motor 183 to rotate the screw nut 182. As a result, the position of the toggle support 130 relative to the tie bar 140 is adjusted, and the distance L between the fixed platen 110 and the toggle support 130 is adjusted. Multiple mold thickness adjustment mechanisms may be used in combination.

[0043] The interval L is detected using the mold thickness adjustment motor encoder 184. The mold thickness adjustment motor encoder 184 detects the amount and direction of rotation of the mold thickness adjustment motor 183 and sends a signal indicating the detection result to the control device 700. The detection result of the mold thickness adjustment motor encoder 184 is used to monitor and control the position and interval L of the toggle support 130. Note that the toggle support position detector for detecting the position of the toggle support 130 and the interval detector for detecting the interval L are not limited to the mold thickness adjustment motor encoder 184, but general-purpose devices can be used.

[0044] The clamping device 100 may have a mold temperature controller that adjusts the temperature of the mold device 800. The mold device 800 has a flow path for a temperature-controlled medium inside it. The mold temperature controller adjusts the temperature of the mold device 800 by adjusting the temperature of the temperature-controlled medium supplied to the flow path of the mold device 800.

[0045] In this embodiment, the mold clamping device 100 is a horizontal type in which the mold opening and closing direction is horizontal, but it may also be a vertical type in which the mold opening and closing direction is vertical.

[0046] In this embodiment, the clamping device 100 has a clamping motor 160 as a drive source, but a hydraulic cylinder may be used instead of the clamping motor 160. Furthermore, the clamping device 100 may have a linear motor for opening and closing the mold, and an electromagnet for clamping the mold.

[0047] (Ejector device) In describing the ejector device 200, similar to the description of the clamping device 100, the direction of movement of the movable platen 120 when the mold is closed (for example, the positive X-axis direction) is described as forward, and the direction of movement of the movable platen 120 when the mold is open (for example, the negative X-axis direction) is described as backward.

[0048] The ejector device 200 is attached to the movable platen 120 and moves back and forth together with the movable platen 120. The ejector device 200 includes an ejector rod 210 that ejects the molded product from the mold device 800 and a drive mechanism 220 that moves the ejector rod 210 in the direction of movement of the movable platen 120 (in the X-axis direction).

[0049] The ejector rod 210 is positioned to move back and forth within a through-hole in the movable platen 120. The front end of the ejector rod 210 contacts the ejector plate 826 of the movable mold 820. The front end of the ejector rod 210 may or may not be connected to the ejector plate 826.

[0050] The drive mechanism 220 includes, for example, an ejector motor and a motion conversion mechanism that converts the rotational motion of the ejector motor into the linear motion of the ejector rod 210. The motion conversion mechanism includes a screw shaft and a screw nut that screws onto the screw shaft. A ball or roller may be interposed between the screw shaft and the screw nut.

[0051] The ejector device 200 performs the ejection process under the control of the control device 700. In the ejection process, the ejector rod 210 is advanced from the standby position to the ejection position at a set travel speed, thereby advancing the ejector plate 826 and ejecting the molded product. Subsequently, the ejector motor is driven to retract the ejector rod 210 at a set travel speed, retracting the ejector plate 826 back to its original standby position.

[0052] The position and speed of the ejector rod 210 are detected, for example, using an ejector motor encoder. The ejector motor encoder detects the rotation of the ejector motor and sends a signal indicating the detection result to the control device 700. Note that the ejector rod position detector, which detects the position of the ejector rod 210, and the ejector rod speed detector, which detects the speed of the ejector rod 210, are not limited to ejector motor encoders, but general-purpose devices can be used.

[0053] (injection device) In the description of the injection device 300, unlike the descriptions of the clamping device 100 and the ejector device 200, the direction of movement of the screw 330 during filling (for example, the negative X-axis direction) is described as forward, and the direction of movement of the screw 330 during metering (for example, the positive X-axis direction) is described as backward.

[0054] The injection device 300 is mounted on a slide base 301, which is positioned to move back and forth relative to the injection device frame 920. The injection device 300 is positioned to move back and forth relative to the mold device 800. The injection device 300 touches the mold device 800 and fills the cavity space 801 in the mold device 800 with the molding material metered in the cylinder 310. The injection device 300 includes, for example, a cylinder 310 for heating the molding material, a nozzle 320 provided at the front end of the cylinder 310, a screw 330 positioned within the cylinder 310 to move back and forth and to rotate, a metering motor 340 for rotating the screw 330, an injection motor 350 for moving the screw 330 back and forth, and a load detector 360 for detecting the load transmitted between the injection motor 350 and the screw 330.

[0055] The cylinder 310 heats the molding material supplied to its interior from the supply port 311. The molding material includes, for example, resin. The molding material is formed, for example, into pellets and supplied to the supply port 311 in a solid state. The supply port 311 is formed at the rear of the cylinder 310. A cooler 312, such as a water-cooled cylinder, is provided on the outer circumference of the rear of the cylinder 310. In front of the cooler 312, a heater 313, such as a band heater, and a temperature detector 314 are provided on the outer circumference of the cylinder 310.

[0056] The cylinder 310 is divided into multiple zones along its axial direction (for example, the X-axis direction). A heater 313 and a temperature detector 314 are provided in each of the multiple zones. A set temperature is set for each of the multiple zones, and the control device 700 controls the heater 313 so that the temperature detected by the temperature detector 314 becomes the set temperature.

[0057] The nozzle 320 is located at the front end of the cylinder 310 and is pressed against the mold device 800. A heater 313 and a temperature detector 314 are provided on the outer circumference of the nozzle 320. The control device 700 controls the heater 313 so that the detected temperature of the nozzle 320 reaches a set temperature.

[0058] The screw 330 is rotatably and reciprocally positioned within the cylinder 310. When the screw 330 is rotated, the molding material is fed forward along the helical groove of the screw 330. As the molding material is fed forward, it is gradually melted by the heat from the cylinder 310. As the liquid molding material is fed forward to the screw 330 and accumulates at the front of the cylinder 310, the screw 330 is retracted. Then, when the screw 330 is advanced, the liquid molding material accumulated in front of the screw 330 is injected from the nozzle 320 and filled into the mold device 800.

[0059] A backflow prevention ring 331 is mounted on the front of the screw 330 so as to be able to move back and forth, acting as a backflow prevention valve to prevent backflow of the molding material from the front to the rear of the screw 330 when the screw 330 is pushed forward.

[0060] When the screw 330 is advanced, the backflow prevention ring 331 is pushed backward by the pressure of the molding material in front of the screw 330, and retracts relative to the screw 330 to a closed position (see Figure 2) that blocks the flow path of the molding material. This prevents the molding material accumulated in front of the screw 330 from flowing backward.

[0061] On the other hand, when the screw 330 is rotated, the backflow prevention ring 331 is pushed forward by the pressure of the molding material being sent forward along the helical groove of the screw 330, and moves relative to the screw 330 to an open position (see Figure 1) that opens the flow path of the molding material. As a result, the molding material is sent forward of the screw 330.

[0062] The backflow prevention ring 331 may be either a co-rotating type that rotates together with the screw 330, or a non-co-rotating type that does not rotate together with the screw 330.

[0063] The injection device 300 may also have a drive source that moves the backflow prevention ring 331 back and forth between an open position and a closed position relative to the screw 330.

[0064] The metering motor 340 rotates the screw 330. The drive source for rotating the screw 330 is not limited to the metering motor 340; for example, a hydraulic pump or the like may also be used.

[0065] The injection motor 350 moves the screw 330 forward and backward. Between the injection motor 350 and the screw 330, there is a motion conversion mechanism that converts the rotational motion of the injection motor 350 into the linear motion of the screw 330. The motion conversion mechanism has, for example, a screw shaft and a screw nut that screws onto the screw shaft. Balls or rollers may be provided between the screw shaft and the screw nut. The drive source for moving the screw 330 forward and backward is not limited to the injection motor 350, but may also be, for example, a hydraulic cylinder.

[0066] The load detector 360 detects the load transmitted between the injection motor 350 and the screw 330. The detected load is converted into pressure by the control device 700. The load detector 360 is installed in the load transmission path between the injection motor 350 and the screw 330 and detects the load acting on the load detector 360.

[0067] The load detector 360 sends a signal of the detected load to the control device 700. The load detected by the load detector 360 is converted into pressure acting between the screw 330 and the molding material, and is used for controlling and monitoring the pressure the screw 330 receives from the molding material, the back pressure on the screw 330, and the pressure acting from the screw 330 on the molding material.

[0068] Furthermore, the pressure detector used to detect the pressure of the molding material is not limited to the load detector 360, but a general-purpose one can be used. For example, a nozzle pressure sensor or an in-mold pressure sensor may be used. The nozzle pressure sensor is installed on the nozzle 320.

[0069] The injection device 300 performs processes such as metering, filling, and holding pressure under the control of the control device 700. The filling and holding pressure processes may be collectively referred to as the injection process.

[0070] In the weighing process, the weighing motor 340 is driven to rotate the screw 330 at a set rotational speed, and the molding material is fed forward along the helical groove of the screw 330. As this occurs, the molding material is gradually melted. As the liquid molding material is fed forward to the screw 330 and accumulates at the front of the cylinder 310, the screw 330 is retracted. The rotational speed of the screw 330 is detected, for example, using a weighing motor encoder 341. The weighing motor encoder 341 detects the rotation of the weighing motor 340 and sends a signal indicating the detection result to the control device 700. Note that the screw rotational speed detector for detecting the rotational speed of the screw 330 is not limited to the weighing motor encoder 341, and a general type can be used.

[0071] In the weighing process, the injection motor 350 may be driven to apply a set back pressure to the screw 330 in order to limit the rapid retraction of the screw 330. The back pressure on the screw 330 is detected, for example, using a load detector 360. The weighing process is completed when the screw 330 has retracted to the weighing completion position and a predetermined amount of molding material has accumulated in front of the screw 330.

[0072] The position and rotational speed of the screw 330 in the metering process are set together as a series of setting conditions. For example, the metering start position, rotational speed switching position, and metering completion position are set. These positions are arranged in this order from front to back and represent the start and end points of the sections in which the rotational speed is set. The rotational speed is set for each section. There may be one or more rotational speed switching positions. The rotational speed switching positions may not be set. In addition, back pressure is set for each section.

[0073] In the filling process, the injection motor 350 is driven to advance the screw 330 at a set speed, filling the cavity space 801 in the mold device 800 with the liquid molding material accumulated in front of the screw 330. The position and speed of the screw 330 are detected, for example, using an injection motor encoder 351. The injection motor encoder 351 detects the rotation of the injection motor 350 and sends a signal indicating the detection result to the control device 700. When the position of the screw 330 reaches the set position, a switchover from the filling process to the holding pressure process (so-called V / P switching) occurs. The position at which the V / P switching occurs is also called the V / P switching position. The set speed of the screw 330 may be changed depending on the position and time of the screw 330.

[0074] The position and movement speed of the screw 330 during the filling process are set together as a series of setting conditions. For example, the filling start position (also called the "injection start position"), the movement speed switching position, and the V / P switching position are set. These positions are arranged in this order from rear to front and represent the start and end points of the sections in which the movement speed is set. The movement speed is set for each section. There may be one or more movement speed switching positions. The movement speed switching positions may not be set at all.

[0075] For each section in which the movement speed of the screw 330 is set, an upper limit is set for the pressure of the screw 330. The pressure of the screw 330 is detected by the load sensor 360. If the pressure of the screw 330 is below the set pressure, the screw 330 moves forward at the set movement speed. On the other hand, if the pressure of the screw 330 exceeds the set pressure, for the purpose of protecting the mold, the screw 330 moves forward at a slower movement speed than the set movement speed so that the pressure of the screw 330 becomes below the set pressure.

[0076] Furthermore, during the filling process, after the screw 330 reaches the V / P switching position, the screw 330 may be temporarily stopped at the V / P switching position, and then the V / P switching may be performed. Immediately before the V / P switching, instead of stopping the screw 330, the screw 330 may be moved forward or backward at a slow speed. In addition, the screw position detector that detects the position of the screw 330 and the screw movement speed detector that detects the movement speed of the screw 330 are not limited to the injection motor encoder 351, but general-purpose ones can be used.

[0077] In the holding pressure process, the injection motor 350 is driven to push the screw 330 forward, maintaining the pressure of the molding material at the front end of the screw 330 (hereinafter also referred to as "holding pressure") at a set pressure, and pushing the molding material remaining in the cylinder 310 toward the mold device 800. This allows for the replenishment of molding material lost due to cooling shrinkage within the mold device 800. The holding pressure is detected, for example, using a load detector 360. The set value of the holding pressure may be changed according to the elapsed time from the start of the holding pressure process. Multiple holding pressures and holding times for maintaining the holding pressure in the holding pressure process may be set, and may be set together as a series of setting conditions.

[0078] During the holding pressure process, the molding material in the cavity space 801 within the mold device 800 is gradually cooled, and upon completion of the holding pressure process, the entrance to the cavity space 801 is sealed with solidified molding material. This state is called a gate seal, and prevents backflow of molding material from the cavity space 801. After the holding pressure process, the cooling process begins. During the cooling process, the molding material in the cavity space 801 is solidified. To shorten the molding cycle time, a metering process may be performed during the cooling process.

[0079] In this embodiment, the injection device 300 is an in-line screw type, but a pre-plasticization type or the like may also be used. In a pre-plasticization injection device, the molding material molten in a plasticizing cylinder is supplied to the injection cylinder, and the molding material is injected from the injection cylinder into the mold device. In the plasticizing cylinder, a screw is arranged to be rotatable but unable to move back and forth, or a screw is arranged to be rotatable and able to move back and forth. On the other hand, a plunger is arranged to be able to move back and forth in the injection cylinder.

[0080] Furthermore, although the injection device 300 in this embodiment is a horizontal type with the axial direction of the cylinder 310 being horizontal, it may also be a vertical type with the axial direction of the cylinder 310 being vertical. The clamping device combined with the vertical injection device 300 may be vertical or horizontal. Similarly, the clamping device combined with the horizontal injection device 300 may be horizontal or vertical.

[0081] (Mobile device) In describing the moving device 400, similar to the description of the injection device 300, the direction of movement of the screw 330 during filling (for example, the negative X-axis direction) is described as forward, and the direction of movement of the screw 330 during metering (for example, the positive X-axis direction) is described as backward.

[0082] The moving device 400 moves the injection device 300 forward and backward relative to the mold device 800. The moving device 400 also presses the nozzle 320 against the mold device 800, generating nozzle touch pressure. The moving device 400 includes a hydraulic pump 410, a motor 420 as a drive source, a hydraulic cylinder 430 as a hydraulic actuator, and the like.

[0083] The hydraulic pump 410 has a first port 411 and a second port 412. The hydraulic pump 410 is a bidirectional rotatable pump, and by switching the rotation direction of the motor 420, it can draw in working fluid (e.g., oil) from either the first port 411 or the second port 412 and discharge it from the other to generate hydraulic pressure. The hydraulic pump 410 can also draw working fluid from a tank and discharge it from either the first port 411 or the second port 412.

[0084] Motor 420 operates the hydraulic pump 410. Motor 420 drives the hydraulic pump 410 with a rotational direction and rotational torque corresponding to the control signal from the control device 700. Motor 420 may be an electric motor or an electric servo motor.

[0085] The hydraulic cylinder 430 comprises a cylinder body 431, a piston 432, and a piston rod 433. The cylinder body 431 is fixed to the injection device 300. The piston 432 divides the inside of the cylinder body 431 into a front chamber 435 as a first chamber and a rear chamber 436 as a second chamber. The piston rod 433 is fixed to the fixed platen 110.

[0086] The front chamber 435 of the hydraulic cylinder 430 is connected to the first port 411 of the hydraulic pump 410 via a first passage 401. The hydraulic fluid discharged from the first port 411 is supplied to the front chamber 435 via the first passage 401, pushing the injection device 300 forward. As the injection device 300 moves forward, the nozzle 320 is pressed against the fixed mold 810. The front chamber 435 functions as a pressure chamber that generates nozzle touch pressure on the nozzle 320 by the pressure of the hydraulic fluid supplied from the hydraulic pump 410.

[0087] Meanwhile, the rear chamber 436 of the hydraulic cylinder 430 is connected to the second port 412 of the hydraulic pump 410 via the second passage 402. The working fluid discharged from the second port 412 is supplied to the rear chamber 436 of the hydraulic cylinder 430 via the second passage 402, pushing the injection device 300 backward. As the injection device 300 is retracted, the nozzle 320 is separated from the fixed mold 810.

[0088] In this embodiment, the moving device 400 includes a hydraulic cylinder 430, but the present invention is not limited thereto. For example, instead of the hydraulic cylinder 430, an electric motor and a motion conversion mechanism that converts the rotational motion of the electric motor into the linear motion of the injection device 300 may be used.

[0089] (Control device) The control device 700 is, for example, a computer and, as shown in Figures 1 and 2, has a CPU (Central Processing Unit) 701, a storage medium 702 such as memory, an input interface 703, an output interface 704, and a communication interface 705. The control device 700 performs various controls by having the CPU 701 execute a program stored in the storage medium 702. The control device 700 also receives signals from the outside through the input interface 703 and transmits signals to the outside through the output interface 704.

[0090] The control device 700 repeatedly produces molded products by repeatedly performing processes such as metering, mold closing, pressure increasing, mold clamping, filling, holding pressure, cooling, depressurization, mold opening, and ejection. A series of operations to obtain a molded product, such as the operations from the start of one metering process to the start of the next metering process, is also called a "shot" or "molding cycle." The time required for one shot is also called the "molding cycle time" or "cycle time."

[0091] A single molding cycle includes, for example, a weighing process, a mold closing process, a pressurizing process, a clamping process, a filling process, a holding pressure process, a cooling process, a depressurizing process, a mold opening process, and an ejection process, in this order. The order here refers to the order in which each process begins. The filling, holding pressure, and cooling processes take place during the clamping process. The start of the clamping process may coincide with the start of the filling process. The completion of the depressurizing process coincides with the start of the mold opening process.

[0092] Furthermore, multiple processes may be performed simultaneously to shorten the molding cycle time. For example, the metering process may be performed during the cooling process of the previous molding cycle, or during the mold clamping process. In this case, the mold closing process may be performed at the beginning of the molding cycle. The filling process may also be started during the mold closing process. The ejection process may also be started during the mold opening process. If an on-off valve is provided to open and close the flow path of the nozzle 320, the mold opening process may be started during the metering process. This is because even if the mold opening process is started during the metering process, if the on-off valve closes the flow path of the nozzle 320, the molding material will not leak from the nozzle 320.

[0093] Furthermore, a single molding cycle may include steps other than the weighing step, mold closing step, pressurization step, mold clamping step, filling step, holding pressure step, cooling step, depressurization step, mold opening step, and ejection step.

[0094] For example, after the holding pressure process is completed and before the metering process begins, a pre-metering suck-back process may be performed in which the screw 330 is retracted to a preset metering start position. This reduces the pressure of the molding material accumulated in front of the screw 330 before the metering process begins and prevents the screw 330 from retracting too quickly at the start of the metering process.

[0095] Furthermore, after the metering process is completed and before the filling process begins, a post-metering suck-back process may be performed in which the screw 330 is retracted to a preset filling start position (also called the "injection start position"). This reduces the pressure of the molding material accumulated in front of the screw 330 before the filling process begins and prevents leakage of the molding material from the nozzle 320 before the filling process begins.

[0096] The control device 700 is connected to an operating device 750 that accepts user input operations and a display device 760 that displays a screen. The operating device 750 and the display device 760 may be integrated, for example, by a touch panel 770. The touch panel 770, as the display device 760, displays a screen under the control of the control device 700. The screen of the touch panel 770 may display information such as the settings of the injection molding machine 10 and the current status of the injection molding machine 10. The touch panel 770 accepts operations in the displayed screen area. The screen area of ​​the touch panel 770 may also display operation parts such as buttons and input fields that accept user input operations. The touch panel 770, as the operating device 750, detects user input operations on the screen and outputs a signal corresponding to the input operation to the control device 700. This allows, for example, the user to operate the operation parts provided on the screen while confirming the information displayed on the screen to configure the injection molding machine 10 (including inputting setting values), etc. Furthermore, the user can operate the injection molding machine 10 corresponding to the operation unit by operating the operation unit provided on the screen. The operation of the injection molding machine 10 may include, for example, the operation (including stopping) of the clamping device 100, ejector device 200, injection device 300, moving device 400, etc. Alternatively, the operation of the injection molding machine 10 may include switching the screen displayed on the touch panel 770, which serves as the display device 760.

[0097] Although the operating device 750 and display device 760 in this embodiment have been described as being integrated as a touch panel 770, they may be provided independently. Furthermore, multiple operating devices 750 may be provided. The operating device 750 and display device 760 are positioned on the operating side (negative Y-axis direction) of the clamping device 100 (more specifically, the fixed platen 110).

[0098] (First Embodiment) There is a demand to perform anomaly detection using a pre-trained model during injection molding. In this case, machine learning is required to generate the trained model.

[0099] It is conceivable to perform the machine learning phase using an injection molding machine. However, this requires high processing power from the injection molding machine's processing unit. Therefore, it is conceivable to perform the learning process using an information processing unit separate from the injection molding machine.

[0100] Therefore, in this embodiment, we will describe an example in which the training phase and the inference phase are performed on different devices. Specifically, the trained model created in the training phase of the machine learning device is mounted on an injection molding machine. Although this embodiment describes an example in which the training phase and the inference phase are performed on different devices, it is not limited to a method in which the training phase and the inference phase are performed on different devices, and the training phase and the inference phase may be performed on the same device.

[0101] Figure 3 is a conceptual diagram illustrating the coordination between the machine learning apparatus and the control device of the injection molding machine according to this embodiment, specifically regarding the trained model.

[0102] The example shown in Figure 3 includes a test injection molding machine 1350, a machine learning device 1300, and a control device 700 for the injection molding machine 10.

[0103] In the example shown in Figure 3, the test injection molding machine 1350 and the machine learning device 1300 may be owned, for example, by the manufacturer of the injection molding machine 10, by the recipient of the injection molding machine, or by a service provider that generates the trained model.

[0104] The test injection molding machine 1350 will be used for machine learning. The configuration of the test injection molding machine 1350 is the same as that of injection molding machine 10, so the explanation will be omitted. If the customer already owns the machine learning device 1300, injection molding machine 10 will be used instead of the test injection molding machine 1350.

[0105] In this embodiment, the test injection molding machine 1350 molds a molded product (an example of a product) according to user settings. The test injection molding machine 1350 then outputs waveform data measured during injection molding to the machine learning device 1300. In this embodiment, the data showing the results detected by the detection device provided in the test injection molding machine 1350 or the injection molding machine 10 in a time series is referred to as waveform data (an example of time series data).

[0106] The machine learning device 1300 stores the input waveform data in the data storage unit 1321.

[0107] The machine learning device 1300 executes the learning phase. To do this, the learning unit 1312 of the machine learning device 1300 loads the data stored in the data storage unit 1321 into the neural network and generates a network with adjusted synaptic weighting and bias as a trained model LM.

[0108] For example, the learning unit 1312 of the machine learning device 1300 generates a trained model LM by reading a large amount of waveform data and performing machine learning using backpropagation with a neural network.

[0109] The machine learning device 1300 may be, for example, an on-premise server installed in a factory or the like, or a cloud server. Furthermore, the machine learning device 1300 may be a stationary terminal device located in a factory or the like, or a portable terminal device (mobile terminal). A stationary terminal device may include, for example, a desktop PC (Personal Computer). A portable terminal device may include, for example, a smartphone, a tablet device, or a laptop PC.

[0110] The control device 700 of the injection molding machine 10 executes an inference phase. The inference unit 714A of the control device 700 inputs waveform data, which represents the results detected by the detection device installed in the injection molding machine 10 in a time series, to the trained model LM and causes the trained model LM to perform inference. Then, the control device 700 detects anomalies based on the output results from the trained model LM.

[0111] In this embodiment, the learning unit 1312 of the machine learning device 1300 generates a trained model LM, and the communication control unit 1313 uses a communication interface to transfer the generated trained model LM to the control device 700. This embodiment is not limited to transferring the trained model LM itself; the weights and biases of the trained model LM may also be transferred to the control device 700. The control device 700 can update the trained model LM based on the weights and biases it receives, thereby matching it with the trained model LM of the learning unit 1312 of the machine learning device 1300.

[0112] As a result, the communication control unit 711 of the control device 700 receives the trained model LM, and the inference unit 714A performs inference using the trained model LM generated in the training phase of the machine learning device 1300. Next, the configurations of the machine learning device 1300 and the control device 700 will be described.

[0113] Figure 4 shows an example of the functional configuration of the machine learning device 1300 according to this embodiment. The functions of the machine learning device 1300 shown in Figure 4 can be realized by any hardware or any combination of hardware and software. For example, as shown in Figure 4, the machine learning device 1300 includes a CPU 1301, a storage medium 1302, and a communication interface 1303.

[0114] The storage medium 1302 stores various installed programs, as well as files and data necessary for various processes. The storage medium 1302 includes, for example, an HDD (Hard Disk Drive), an SSD (Solid State Drive), or flash memory.

[0115] The storage medium 1302 according to this embodiment comprises a data storage unit 1321 and a trained model LM. The data storage unit 1321 stores training data used for training the trained model LM.

[0116] The communication interface 1303 is used as an interface for connecting to external devices in a communicative manner. This allows the machine learning device 1300 to communicate with external devices, such as an injection molding machine 10, through the communication interface 1303. Furthermore, the communication interface 1303 may have multiple types of communication interfaces depending on the communication method between it and the connected devices.

[0117] The CPU 1301 of the machine learning device 1300 executes a program stored in the storage medium 1302. The CPU 1301 includes, as a functional unit, an acquisition unit 1311, a learning unit 1312, and a communication control unit 1313.

[0118] First, we will describe the configuration of the machine learning device 1300 for generating the trained model LM.

[0119] The acquisition unit 1311 acquires waveform data from the test injection molding machine 1350, showing the detection results from a detection device installed in the test injection molding machine 1350, and stores the acquired waveform data in the data storage unit 1321. The detection results acquired from the detection device are, for example, time-series data of the pressure received from the molding material, detected by the load detector 360. However, this embodiment does not limit the detection results acquired from the detection device to time-series data of the pressure received from the molding material; time-series data of other physical quantities such as position, speed, or torque of the components included in the test injection molding machine 1350 may also be used. Furthermore, time-series data of detection results from a detection device installed inside or outside the test injection molding machine 1350 may also be used. Moreover, this embodiment does not limit the use of one type of time-series data; multiple types of time-series data may be used. For example, by using time-series data of the torque generated when the movable platen 120 is moved during mold opening and closing, it is possible to detect abnormalities related to foreign matter contamination in the mold device 800 or the mold clamping device 100.

[0120] Figure 5 shows an example of waveform data of the pressure value (hereinafter referred to as the actual pressure value) detected by the load detector 360 installed in the test injection molding machine 1350 or injection molding machine 10 according to this embodiment. The example shown in Figure 5 shows the time-series change of the actual pressure value and the shot number acquired by the built-in load detector 360 during molding by the test injection molding machine 1350 or injection molding machine 10. In the example shown in Figure 5, the shot number increases at times t1, t2, and t3. The example shown in Figure 5 shows the case where the detection period is every second, but an appropriate period can be set according to the embodiment.

[0121] Figure 6 is a diagram illustrating, in a table structure, the actual pressure values ​​detected by a load detector 360 installed in the test injection molding machine 1350 or injection molding machine 10 according to this embodiment. In the example shown in Figure 6, the time, shot number, elapsed time from the trigger, and actual pressure values ​​are shown in correspondence. The trigger is set to the timing of a switching of the molding process, such as the start of mold closing or the start of injection. In other words, the elapsed time from the trigger according to this embodiment is the elapsed time from the timing of a switching of the molding process, such as the start of mold closing or the start of injection, but it may be set to the time from the switching of any process, or even the elapsed time from when the shot number changes. The actual pressure value is the actual pressure value acquired (measured) by the load detector 360. In this embodiment, the test injection molding machine 1350 and the injection molding machine 10 manage the switching of processes. Therefore, the acquisition unit 1311 according to this embodiment grasps the process change through communication from the test injection molding machine 1350 or the injection molding machine 10, and collects information using the process change as a trigger. Thus, in this embodiment, it is possible to synchronize the timing of information collection.

[0122] For example, the acquisition unit 1311 acquires the detection result from a detection device such as a load detector 360 as a pressure actual value for each shot, starting from the trigger timing, and stores it in the data storage unit 1321 in association with the time, shot number, and elapsed time since the trigger. Alternatively, the acquisition unit 1311 may constantly acquire sensor signals and store the data detected at process switching (e.g., pressure actual value) in association with the time, shot number, and elapsed time since the trigger. In the test injection molding machine 1350 and injection molding machine 10 according to this embodiment, the detection results from a detection device such as a load detector 360 are buffered. The buffered data is not limited to data after the trigger timing, but also includes data before the trigger timing. For example, in the test injection molding machine 1350 and injection molding machine 10, when buffering 550 data points as a waveform function, 50 data points are stored before the trigger timing and 500 data points are stored after the trigger timing. In other words, if 100% of the data to be buffered is taken from the trigger timing onward, the test injection molding machine 1350 and the injection molding machine 10 buffer only 10% of the data from before the trigger timing. Thus, although this embodiment acquires data starting from the trigger, the trigger timing is not limited to the starting position of the waveform data.

[0123] The acquisition unit 1311 acquires waveform data (an example of time-series data) showing the pressure detected by the load detector 360 from the test injection molding machine 1350 each time a molded product is formed. In this embodiment, the acquired waveform data is not limited to waveform data showing the pressure detected by the load detector 360 from the start to the end of the molding cycle, but any waveform data showing the detection result for each shot is acceptable. For example, the acquired waveform data may show part or all of the molding cycle, or even waveform data that overlaps with the preceding and succeeding shots.

[0124] In this embodiment, the waveform data stored in the data storage unit 1321 by the acquisition unit 1311 is the waveform data detected by the detection device (e.g., load detector 360) when a normal molded product is formed in the test injection molding machine 1350 or injection molding machine 10. In other words, in this embodiment, the waveform data stored in the data storage unit 1321 does not include waveform data when an abnormality occurs.

[0125] Figures 5 and 6 show that the cycle time for each shot used to mold the product is 10 to 11 seconds. In this embodiment, for each shot, the waveform data showing the results detected by the detection device in time series is input to the input layer of the trained model LM.

[0126] The trained model LM according to this embodiment is an encoder-decoder model having an encoder and a decoder, based on an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layers. The encoder implements the function of extracting important features from input time-series data, and the decoder implements the function of generating time-series data based on the features extracted by the encoder.

[0127] This embodiment describes an example in which a trained model LM functions as an autoencoder. However, this embodiment is not limited to an example in which a trained model LM functions as an autoencoder; any machine learning model that has been trained to estimate time-series data of the results detected by the detection device for each shot in the injection molding machine 10 may be used. The trained model LM according to this embodiment functions to output waveform data that matches the input waveform data when normal waveform data is input. For example, when waveform data of pressure for each shot is input, the trained model LM outputs waveform data inferred as time-series data of the results detected by the detection device for the time period corresponding to the input waveform data. This embodiment is not limited to an example in which a machine learning model that functions as an autoencoder is used; for example, a convolutional autoencoder, a VAE (Variational Autoencoder), a GAN (Generative Adversarial Network), a flow-based model, a diffusion model, or other generative model may be used.

[0128] In this embodiment, the data input to the input layer of the trained model LM and the data output from the output layer have a fixed data size. For example, the trained model LM takes one shot and inputs waveform data of 10 seconds (10 points) of actual pressure values, and outputs waveform data of 10 seconds (10 points) of actual pressure values ​​as the inference result of one shot. Although this embodiment shows an example of inputting 10 points of actual pressure values ​​as an example of a fixed length, this embodiment does not limit the data size, and it may be 9 points or less or 11 points or more. In the machine learning device 1300 according to this embodiment, the triggers for collecting data for machine learning are aligned, and the data size used for machine learning is fixed, so the data for each shot for machine learning the trained model LM can be aligned. Therefore, the machine learning device 1300 according to this embodiment can improve the learning accuracy for generating the trained model LM.

[0129] The learning unit 1312 comprises an inference unit 1312A, an error calculation unit 1312B, and an update unit 1312C. When waveform data stored in the data storage unit 1321 is input, the learning unit 1312 generates a trained model LM by performing machine learning using the waveform data stored in the data storage unit 1321 so as to output waveform data (an example of second time series data) identical to the waveform data stored in the data storage unit 1321 (an example of first time series data). The waveform data stored in the data storage unit 1321 is the waveform data for one shot when a molded product is successfully molded by the test injection molding machine 1350 or the injection molding machine 10. The learning unit 1312 stores the generated trained model LM in the storage medium 1302.

[0130] Figure 7 is a diagram illustrating the concept of machine learning by the learning unit 1312 according to this embodiment. As shown in Figure 7, the inference unit 1312A inputs waveform data 1801 stored in the data storage unit 1321 to a neural network having an encoder and a decoder, and receives waveform data 1802 from the neural network having the same data size (data length) as the input waveform data. The neural network having an encoder and a decoder serves as the basis for the trained model LM.

[0131] The inference unit 1312A uses waveform data 1801 containing 10 seconds (10 points) of pressure data as input to the neural network. Therefore, if 11 seconds (11 points) of pressure data are detected as one shot, the inference unit 1312A removes the last second (1 point) of pressure data and inputs the waveform data of 10 points of pressure data to the neural network.

[0132] The error calculation unit 1312B then calculates the error between the input waveform data 1801 and the output waveform data 1802. The error calculation method may be a well-known method such as the root mean squared error (RMSE), mean squared error (MSE), mean absolute error (MAE), or cosine similarity, and in this embodiment, MSE (mean squared error) is used.

[0133] The update unit 1312C updates the neural network parameters based on the error calculated by the error calculation unit 1312B. For example, the update unit 1312C updates the neural network parameters so that the output waveform data 1802 matches the input waveform data 1801. The parameter update method may be a well-known method, such as backpropagation.

[0134] The learning unit 1312 generates a trained model LM by repeating the process shown in Figure 7 for each waveform data stored in the data storage unit 1321, and stores it in the storage medium 1302.

[0135] As described above, the learning unit 1312 in this embodiment performs machine learning by calculating the amount of change in the parameters related to each node in the neural network from the error between the output value (waveform data, which is the inference result) and the input value (waveform data) of one shot, and updating them. In this embodiment, the data used to compare the output value (waveform data, which is the inference result) used to calculate the error is the input waveform data. The trained model LM thus generated will be a model that, when one shot of waveform data is input, outputs the same waveform data as the input if the waveform data is normal. In this embodiment, by using such a learning method, it is not necessary to prepare correct values, and machine learning can be performed using only the time-series data, which is the detection result of the detection device.

[0136] Furthermore, the trained model LM may be updated by adding a new set of waveform data to the existing trained model LM for further training.

[0137] Furthermore, the learning unit 1312 is provided with a preprocessing unit (not shown) that performs standardization or normalization on the waveform data before the inference unit 1312A performs inference. In addition to standardization or normalization, the preprocessing unit may also perform one or more of the following: missing value processing and resampling. In this embodiment, for example, the preprocessing unit may divide the actual pressure values ​​used in the waveform data by a predetermined pressure value so that the actual pressure values ​​fall within the range of 1.0.

[0138] Returning to Figure 4, the communication control unit 1313 uses the communication interface 1303 to send and receive information with external devices such as the injection molding machine 10. For example, the communication control unit 1313 may transmit the trained model LM stored in the storage medium 1302 to the control device 700 of the injection molding machine 10. Alternatively, the communication control unit 1313 may extract a structure showing the parameters (e.g., weights and biases) set for each layer constituting the trained model LM and transmit this structure to the control device 700.

[0139] In this way, the machine learning device 1300 prepares a set of waveform data based on various molding conditions provided in the test injection molding machine 1350 or injection molding machine 10, and various mold devices provided in the test injection molding machine 1350 or injection molding machine 10. The machine learning device 1300 performs machine learning using this set of waveform data to generate a trained model LM. As a result, inference can be performed using the trained model LM regardless of the molding conditions and the type of mold device 800.

[0140] Therefore, the control device 700 of the injection molding machine 10 can perform abnormality detection using the learned model LM, regardless of the provided mold device 800 and the set molding conditions.

[0141] Figure 8 shows an example of the functional configuration of the control device 700 of the injection molding machine 10 according to this embodiment. Each functional block shown in Figure 8 is conceptual and does not necessarily have to be physically configured as shown. All or part of each functional block can be configured by functionally or physically distributing and integrating them in any unit. Each processing function performed in each functional block is realized in whole or in any part by a program executed by the CPU 701. Alternatively, each functional block may be realized as hardware using wired logic. As shown in Figure 8, the CPU 701 of the control device 700 includes a communication control unit 711, an injection molding processing unit 712, an acquisition unit 713, an anomaly detection unit 714, and a display control unit 715. The control device 700 also has a trained model LM in the storage medium 702. The trained model LM has the same parameters as the trained model LM stored in the storage medium 1302 of the machine learning device 1300.

[0142] The communication control unit 711 transmits and receives information to and from external devices such as the machine learning device 1300 using the communication interface 705. For example, the communication control unit 711 may receive a trained model LM from the machine learning device 1300. The communication control unit 711 may also receive information indicating a structure that shows the parameters (e.g., weights and biases) set for each layer constituting the trained model LM.

[0143] In this embodiment, an example of receiving information indicating a trained model LM or structure from the machine learning device 1300 was described as a method for acquiring such information. However, this does not limit the method for acquiring information indicating a trained model LM or structure. For example, information indicating a trained model LM or structure may be acquired via an external storage medium.

[0144] When the communication control unit 711 receives information indicating the structure of a learned model LM, it may update the learned model LM stored in the storage medium 702.

[0145] The injection molding processing unit 712 performs processing to form a molded product in the injection molding machine 10. For example, when forming a molded product, the injection molding processing unit 712 may set each item that constitutes the molding conditions and then perform injection molding.

[0146] The acquisition unit 713 acquires waveform data (an example of first time-series data) showing the detection result from a detection device provided in the injection molding machine 10 in time series for each shot in which a molded product is formed by the injection molding machine 10. In this embodiment, the waveform data to be acquired is an example in which time-series data showing the pressure received from the molding material in time series, detected by a load detector 360, which is an example of a detection device. In this embodiment, the detection result acquired from the detection device to detect an abnormality is not limited to time-series data of the pressure received from the molding material, but may also be other physical quantities such as the position, speed, or torque of the components included in the injection molding machine 10. Furthermore, time-series data of the detection result from a detection device provided inside or outside the injection molding machine 10 may also be used. In addition, this embodiment is not limited to the use of one type of time-series data, but may also use multiple types of time-series data.

[0147] The anomaly detection unit 714 includes an inference unit 714A, and performs an anomaly detection process in the injection molding machine 10 based on the waveform data acquired by the acquisition unit 713.

[0148] The inference unit 714A inputs waveform data, which represents the results detected by a detection device (e.g., a load detector 360) in a time series, to the input layer of the trained model LM while the molded product is being formed by the injection molding processing unit 712. The inference unit 714A receives from the output layer of the trained model LM waveform data (an example of output data) that has been inferred as time series data of the results detected by the detection device during the time period corresponding to the input waveform data. In this embodiment, the pressure value included in the waveform data output from the trained model LM is referred to as the pressure inference value.

[0149] Furthermore, if the waveform data acquired by the acquisition unit 713 is larger than the fixed-length data size of the input layer of the trained model LM, the inference unit 714A deletes data (actual pressure values ​​detected after 10 seconds (example of a predetermined cycle time)) from the waveform data so that it becomes the fixed-length data size, and inputs the waveform data, which has the same fixed-length data size as the input layer, to the trained model LM. In this embodiment, the inference unit 714A can reduce the computational load compared to the case of resampling by fixing the waveform data length through the above-described process.

[0150] Furthermore, the anomaly detection unit 714 is provided with a preprocessing unit (not shown) that performs standardization or normalization on the waveform data before inference is performed by the inference unit 714A. In addition to standardization or normalization, the preprocessing unit may also perform one or more of the following: missing value processing and resampling. In this embodiment, for example, the preprocessing unit divides the actual pressure values ​​included in the waveform data acquired by the acquisition unit 713 by a predetermined pressure value to perform preprocessing so that the actual pressure values ​​fall within the range of 1.0. In this embodiment, by performing preprocessing, the estimation accuracy by the trained model LM is improved, thereby improving the accuracy of anomaly detection.

[0151] The anomaly detection unit 714 then performs control to detect anomalies based on the difference between the waveform data input to the trained model LM and the waveform data output from the trained model LM, which is the inference result.

[0152] The display control unit 715 displays information on the display device 760. For example, the display control unit 715 displays the abnormality detection result from the abnormality detection unit 714.

[0153] Figure 9 is a diagram illustrating the concept of the processing performed by the anomaly detection unit 714 and the display control unit 715 according to this embodiment. As shown in Figure 9, the inference unit 714A provided in the anomaly detection unit 714 inputs waveform data 1901 acquired by the acquisition unit 713 to a trained model LM having an encoder and a decoder, and receives waveform data 1902 from the trained model LM having the same data size (data length) as the input waveform data.

[0154] The inference unit 714A uses waveform data 1901 containing 10 seconds (10 points) of pressure data as input to the trained model LM. Therefore, if 11 seconds (11 points) of pressure data are detected as one shot, the inference unit 714A removes the last second (1 point) of pressure data and inputs the waveform data of 10 points of pressure data to the trained model LM.

[0155] The waveform data 1901 shown in Figure 9 represents an example where an anomaly occurs in the actual pressure value 1901A. Therefore, the waveform data 1902 output from the trained model LM has been corrected for the actual pressure value 1901A.

[0156] The anomaly detection unit 714 calculates a value (hereinafter also referred to as the pressure anomaly value) that indicates the degree of anomaly between the input waveform data 1901 and the output waveform data 1902. The method for calculating the value indicating the degree of anomaly may be a well-known method, such as the root mean squared error (RMSE), mean squared error (MSE), mean absolute error (MAE), or cosine similarity. In this embodiment, MSE (mean squared error) is used.

[0157] Figure 10 is an explanatory diagram showing the residual calculated by the anomaly detection unit 714 according to this embodiment. In the example shown in Figure 10, the time, shot number, elapsed time from the trigger, actual pressure value, estimated pressure value, and residual are shown. In the example shown in Figure 10, the actual pressure value and the estimated pressure value are assumed to be normalized to 0-100% by preprocessing.

[0158] In the example shown in Figure 10, the actual pressure value contained in the input waveform data 1901 and the estimated pressure value contained in the output waveform data 1902 are shown for each elapsed time from the trigger at shot number "1004". Furthermore, the residual obtained by subtracting the actual pressure value from the estimated pressure value is also shown.

[0159] In the example shown in Figure 10, the residual is "23" at a time of "7" elapsed since the trigger of shot number "1004". This elapsed time of "7" from the trigger is the time when the actual pressure value 1901A of waveform data 1901 in Figure 9 occurred.

[0160] Figure 11 is a graph showing the actual pressure, estimated pressure, and residual for shot number "1004" using the injection molding machine 10 according to this embodiment. In the example shown in Figure 11, line 2101 shows the actual pressure, line 2102 shows the estimated pressure, and line 2103 shows the residual.

[0161] As shown in Figure 11, at a time interval of "7" from the trigger, a discrepancy occurs between the actual pressure value and the estimated pressure value, resulting in a large residual.

[0162] As described above, the inference performed by the inference unit 714A will produce an inference result that reproduces the input waveform data as closely as possible, provided that the molding of the molded product is performed correctly.

[0163] In contrast, in the examples shown in Figures 10 and 11, the residual between the pressure inferred value and the pressure actual value is large at the elapsed time of "7" from the trigger of shot number "1004".

[0164] In the case of a trained model LM that functions as an autoencoder, as in this embodiment, there is a tendency for it to be unable to correctly infer waveform data that has abnormal features that deviate from the trends of the data used for training.

[0165] Therefore, the abnormality detection unit 714 according to the present embodiment evaluates the error between the input waveform data and the output waveform data by utilizing the property that waveform data with abnormal features cannot be correctly inferred, thereby detecting whether an abnormality occurs for each shot.

[0166] For example, the abnormality detection unit 714 uses the MSE (Mean Squared Error) to detect whether an abnormality occurs. In the present embodiment, the pressure abnormality degree is used as an index for determining whether an abnormality occurs. In the case of the residual [-1, 0, -1, 1, 0, 0, 0, 23, 0, -1] (unit: “%”) of the shot number “1004”, the abnormality detection unit 714 calculates the pressure abnormality degree from the following formula (1). In the example shown by formula (1), “0.533 (%)” is calculated as the pressure abnormality degree.

[0167] {(-0.01) 2 +(0) 2 +(-0.01) 2 +(0.01) 2 +(0) 2 +(0) 2 +(0) 2 +(0.23) 2 +(0) 2 +(-0.01) 2} / 10 = 0.00533……(1)

[0168] For example, the abnormality detection unit 714 can determine whether an abnormality occurs according to whether the calculated pressure abnormality degree is greater than or equal to the threshold value. The threshold value may be determined according to the implementation mode.

[0169] The pressure abnormality degree calculated by the MSE (Mean Squared Error) etc. reflects the degree of abnormality of the waveform data. And in the present embodiment, the calculated pressure abnormality degree may be monitored by the same method as the conventional one.

[0170] The display control unit 715 displays the result calculated by the abnormality detection unit 714. FIG. 12 is a diagram showing an example of the screen displayed by the display control unit 715 according to the present embodiment on the display device 760.

[0171] The example screen shown in Figure 12 is a screen for displaying information for each shot as a graph. For example, in screen 2200, the input field 2201 for the X axis is set to "10s (seconds)", the input field 2202 for the Y axis is set to "Ratio", and the trigger 2203 for starting the display is set to "Type Closure Start". According to these settings, the graph is displayed in the display area 2220.

[0172] In the screen 2200 shown in Figure 12, when the "cursor" button 2204 or the "grid" button 2205 is pressed, the display control unit 715 switches the display mode of the graph displayed in the display area 2220.

[0173] In the screen 2200 shown in Figure 12, when the "Overwrite" button 2206 is pressed, the display control unit 715 overlays and displays the graph for each shot in the display area 2220.

[0174] In the screen 2200 shown in Figure 12, when the "1 shot save" button 2207 is pressed, the control device 700 performs the necessary actions to save the information of the current shot displayed in the display area 2220. When the "CLEAR" button 2208 is pressed, the control device 700 initializes the display in the display area 2220.

[0175] As shown in Figure 12, the information set for "CH-1" to "CH-3" on screen 2200 is displayed in the display area 2220.

[0176] The setting field 2211A for "CH-1" is set to "Actual Pressure Value". Furthermore, the lower limit setting field 2211B for "CH-1" is set to "-50", and the upper limit setting field 2211C is set to "150". Since setting field 2211D is set to "ON", the information set for "CH-1" is displayed in the display area 2220.

[0177] As a result, the display control unit 715 displays the waveform data for one shot of the actual pressure value received from the molding material, detected by the load detector 360, as line 2221 in the display area 2220. The upper limit of the Y axis in the display area 2220 is "150", and the lower limit is "-50". Since it is a ratio display, the unit is [%].

[0178] The setting field 2212A for "CH-2" is set to "Estimated Pressure". Furthermore, the lower limit setting field 2212B for "CH-2" is set to "-50", and the upper limit setting field 2212C is set to "150". Since setting field 2212D is set to "On", the information set for "CH-2" is displayed in the display area 2220.

[0179] As a result, the display control unit 715 displays the waveform data showing the pressure estimate for one shot (10 seconds), output by the inference unit 714A, as line 2222 in the display area 2220. The upper limit of the Y axis in the display area 2220 is "150", and the lower limit is "-50". Since it is a ratio display, the unit is [%].

[0180] In the setting field 2213A for "CH-3", "Pressure Residual Difference" is set. Furthermore, in the lower limit setting field 2213B for "CH-3", "-50" is set, and in the upper limit setting field 2213C, "150" is set. Since "ON" is set in setting field 2213D, the information set for "CH-3" is displayed in the display area 2220.

[0181] As a result, the display control unit 715 displays the change in the residual obtained by subtracting the actual pressure value from the estimated pressure value as line 2223 in the display area 2220. The upper limit of the Y axis in the display area 2220 is "150", and the lower limit is "-50". Since it is a ratio display, the unit is [%].

[0182] In this way, the display device 760, under the control of the display control unit 715, inputs waveform data representing the time-series results detected by a detection device provided on the injection molding machine 10 during one shot in which a molded product is formed by the injection molding machine 10. The difference (residual) between the waveform data received as an inference result from the trained model LM and the said waveform data is represented as a graph and displayed on the display panel (an example of a display unit).

[0183] By referring to the screen 2200 shown in Figure 12, the user can recognize whether or not an abnormality has occurred based on whether or not the pressure residual is excited, etc. Thus, in the example shown in Figure 12, the user can understand whether or not an abnormality has occurred by displaying at least one of the pressure inference value and pressure residual as waveforms. Furthermore, the display device 760 may also use a method of displaying a threshold value for the pressure residual on the screen 2200 and detecting an abnormality depending on whether or not it is above the threshold value.

[0184] The screen examples displayed by the display control unit 715 according to this embodiment are not limited to the screen example shown in Figure 12.

[0185] Figure 13 is an example of a log information screen output by the display control unit 715 according to this embodiment.

[0186] The log information screen 2300 shown in Figure 13 displays the monitoring settings section 2301, the monitoring range settings section 2302, the statistics list 2320, and the performance list 2330.

[0187] The statistics list 2320 displays the statistical values ​​(e.g., mean, range, maximum, minimum, standard deviation) for each setting field 2321 to 2323. The contents displayed in setting fields 2321 to 2323 can be configured by the user. In this embodiment, the items displayed in setting fields 2321 to 2323 can be displayed and monitored. In this embodiment, monitoring refers to the determination of whether or not a product is good based on predetermined criteria.

[0188] The "Monitoring Value," "Width+," and "Width-" values ​​in Statistics List 2320 are used to determine whether or not the molded product is defective in the corresponding setting field.

[0189] This indicates that the control device 700 performs monitoring when the monitoring setting in the statistics list 2320 is set to "fixed". When it is set to "fixed", the anomaly detection unit 714 of the control device 700 determines whether the measured actual value for the item indicated in the setting field meets the criteria indicated by "monitoring value", "width +", and "width -".

[0190] In Statistical List 2320, "Defective" indicates the number of molded parts that do not meet the criteria shown by "Monitoring Value," "Width +," and "Width -."

[0191] The "Cycle Time" setting in setting field 2321, the "Total Peak Pressure" setting in setting field 2322, and the "Pressure Anomaly" setting in setting field 2323 are items set to monitor cycle, filling, and pressure anomalies.

[0192] The "Pressure Anomaly Degree" displays the pressure anomaly degree calculated by the anomaly detection unit 714.

[0193] Note that settings fields 2321-2323 can be changed to items that the user wants to monitor. The method for making these changes will not be explained. The log information screen 2300 shown in Figure 13 does not limit the number of items shown in the statistics list 2320 to three; for example, four or more items may be displayed.

[0194] The results list 2330 shows, for each shot, a list of actual values ​​measured by various detection devices or inference results inferred by a machine learning model (e.g., a trained model LM) for items set in settings fields 2321 to 2323. Items set in settings fields 2321 to 2323 are set to "CH-1" through "CH-3".

[0195] Furthermore, the performance list 2330 associates information about each shot, such as the "shot number," the "time" at which injection molding was performed, and the "classification," which is the result of monitoring set by the statistics list 2320. In addition, the performance list 2330 displays "Inferring" while inference is being performed by a machine learning model (e.g., the trained model LM). After the inference using the machine learning model (e.g., the trained model LM) by the inference unit 714A is completed and the value to be displayed in that item is calculated, the display control unit 715 updates the display to show the calculated value.

[0196] The monitoring setting field 2301 is a pull-down menu that accepts whether or not to monitor according to the items to be monitored in the statistics list 2320. When "On" is selected in the monitoring setting field 2301, the system monitors whether or not an item is defective for each shot, and the monitoring result is displayed in the "Determination" field. The monitoring setting field 2301 switches between "Off" and "On" depending on the user's selection.

[0197] The monitoring range setting field 2302 is a pull-down menu for setting the monitoring range. If "+~-" is selected in the monitoring range setting field 2302, the anomaly detection unit 714 determines whether the calculated value is included within the range indicated by "width+" and "width-" based on the "monitoring value" in the statistics list 2320. For example, if the "monitoring value" is "0", "width+" is "0.1", and "width-" is "0", the anomaly detection unit 714 determines that it is normal if the calculated value is within the range of lower limit "0" ("monitoring value" - "width-") to upper limit "0.1" ("monitoring value" + "width+"), and determines that it is abnormal if it is outside the range of "0" to "0.1".

[0198] For example, in line 2331, the "pressure anomaly" is "0.533," which falls outside the monitoring range of "0" to "0.1." Therefore, the judgment "E (Error)" is displayed.

[0199] In this way, the display device 760, under the control of the display control unit 715, inputs waveform data showing the results detected by a detection device installed in the injection molding machine 10 in a time series during one shot in which the molded product is formed by the injection molding machine 10. The difference between the waveform data received as an inference result from the learned model LM and the waveform data showing the results detected by the detection device in a time series is displayed as an "abnormal pressure value" on the display panel (an example of a display unit).

[0200] The user can refer to the "abnormal pressure value" for each shot on the log information screen 2300 shown in Figure 13 to determine whether or not an abnormality occurred for each shot.

[0201] Figure 14 is a diagram illustrating another aspect of the log information screen output by the display control unit 715 according to this embodiment.

[0202] The log information screen 2400 shown in Figure 14 displays the monitoring settings section 2301, the monitoring range settings section 2302, the statistics list 2320, and the performance list 2430. The log information screen 2400 shown in Figure 14 differs from the log information screen 2300 shown in Figure 13 in the display of the performance list 2430. Other items are the same as those in the log information screen 2300 shown in Figure 13, and their explanation is omitted.

[0203] The results list 2430 displays, for each shot, the actual values ​​measured by various detection devices for the items set in settings fields 2321 to 2323, or the inference results inferred by a machine learning model (e.g., a trained model LM), as a graph. The items set in settings fields 2321 to 2323 are set to "CH-1" through "CH-3". In the results list 2430, the changes in each value from "CH-1" to "CH-3" are displayed in a graph, making it easy for the user to understand the changes for each shot.

[0204] This embodiment describes an example of detecting whether an anomaly has occurred by calculating the difference between the input waveform data and the waveform data output from the trained model LM using MSE (Mean Squared Error). However, this embodiment is not limited to this method, and for example, whether an anomaly has occurred may be detected based on whether the residual (e.g., pressure residual) is greater than or equal to a threshold. This embodiment is merely an example of anomaly detection and is not limited to the detection method described above. Any method that can evaluate the difference between the input waveform data and the waveform data output from the trained model LM for each shot may be used as the method for detecting anomalies.

[0205] (Second embodiment) In the above-described embodiment, an example was explained in which the control device 700 of the injection molding machine 10, acting as a management device for the injection molding machine 10, uses a trained model LM to detect abnormalities. However, the above-described embodiment is not limited to a method in which the control device 700 of the injection molding machine 10 uses a trained model LM to detect abnormalities. Therefore, in the second embodiment, an example will be described in which a group control device 2500 that controls the injection molding machine 10 uses a trained model LM to detect abnormalities.

[0206] Figure 15 is a diagram illustrating the configuration of the machine learning device 1300, the group control device 2500, and the injection molding machine 10 according to this embodiment. As shown in Figure 13, the group control device 2500 manages, for example, eight injection molding machines 10. Note that the number of injection molding machines to be managed is just an example and can be any number.

[0207] This embodiment applies the detection of abnormalities using a learned model LM by the control device 700 described in the above-described embodiment to a group control device 2500 that has a group control function for multiple injection molding machines 10.

[0208] The machine learning device 1300 according to this embodiment has the same configuration as the first embodiment. The machine learning device 1300 transmits information about the trained model LM to the group management device 2500 via the communication line NW.

[0209] The communication line NW is, for example, an internet communication line. When communication is performed between the machine learning device 1300 and the group management device 2500, it is preferable to connect using a VPN (Virtual Private Network). Connecting via VPN can improve the security of the communication.

[0210] The group control device (an example of a control device) 2500 is a device that manages multiple injection molding machines 10 in terms of productivity. It is connected to each injection molding machine 10 and, similar to the control device 700 described above, receives molding conditions and detection results obtained from various detection devices, assisting in the management and planning of production status.

[0211] The group control device 2500 may be implemented as, for example, a personal computer. However, the group control device 2500 does not normally have a control function for the injection molding operation of each injection molding machine 10, but this control function can be added by extending the capabilities of the personal computer.

[0212] The group control device 2500, like the control device 700, has a storage medium (not shown) on which the learned model LM is stored.

[0213] The group management device 2500, like the control device 700, has a CPU (not shown), and by having the CPU execute a program stored in a storage medium, the communication control unit 711, acquisition unit 713, abnormality detection unit 714, and display control unit 715 are realized, just like in the control device 700. The processing performed by each component is the same as in the embodiment described above, so its explanation is omitted.

[0214] The group management device 2500 is made connectable to the machine learning device 1300 via a communication line NW.

[0215] The group management device 2500 receives information about the trained model LM from the machine learning device 1300. Based on the received information, the group management device 2500 updates the trained model LM, etc.

[0216] For example, the acquisition unit 713 of the group control device 2500 acquires waveform data transmitted from each of the injection molding machines 10. Then, the abnormality detection unit 714 of the group control device 2500 determines whether or not an abnormality has occurred based on the acquired waveform data. The determination method is the same as in the embodiment described above, so its explanation is omitted.

[0217] A display device (not shown) is connected to the group control device 2500. The display device connected to the group control device 2500, under the control of the group control device 2500, inputs waveform data showing the results detected by a detection device provided on the injection molding machine 10 in a time series during one shot in which a molded product is formed by the injection molding machine 10. The display panel (an example of a display unit) then displays information showing the difference between the waveform data received as an inference result from the group control device LM and the waveform data in question. As a result, the display device, for example, displays the screens shown in Figures 12 to 14.

[0218] <effect> In the embodiment described above, it is possible to easily detect anomalies based on the difference between waveform data showing the results detected by the detection device in a time series during one shot and waveform data output from the trained model LM.

[0219] Unlike conventional monitoring functions that use detected values ​​at a predetermined point in time, or statistical measures such as maximum or minimum values, this system allows users to easily monitor for anomalies without needing to be familiar with the characteristics of the detected values.

[0220] Conventionally, it was difficult to detect anomalies unless changes were recognized when an anomaly occurred and thresholds were set accordingly. In this embodiment, changes when an anomaly occurs can be easily recognized from the difference between the waveform data output from the trained model LM and the input waveform data, thereby reducing the burden of setting up anomaly detection.

[0221] Another possible method for detecting defects in molded products is to input images of the molded products into a machine learning model. However, detecting quality defects using such images would require the addition of peripheral equipment such as image inspection devices.

[0222] In contrast, in the above-described embodiment, abnormalities can be detected by the detection results of the detection device incorporated into the injection molding machine 10, making it easy to detect quality defects.

[0223] Furthermore, another method of inputting waveform data into a machine learning model is to input image data representing the waveform. In this case, image processing is performed, which increases the computational load.

[0224] In contrast, in the above-described embodiment, waveform data representing the detection results of the detection device in time series, in other words, one-dimensional data, is input to the trained model LM. Therefore, the amount of data input to the trained model LM can be reduced. Consequently, the above-described embodiment can reduce the size of the trained model LM and reduce the processing burden.

[0225] Preferred embodiments of the present disclosure have been described above. However, the inventions of the present disclosure are not limited to the embodiments described above. Various modifications, substitutions, etc., can be applied to the embodiments described above without departing from the scope of the inventions of the present disclosure. Furthermore, each of the features described with reference to the embodiments described above may be combined as appropriate, as long as they do not contradict each other technically. [Explanation of symbols]

[0226] 10 injection molding machine 700 Control Unit 701 CPU 702 Storage medium 711 Communication Control Unit 712 Injection Molding Processing Unit 713 Acquisition Department 714 Anomaly detection unit 714A Reasoning Department 715 Display Control Unit LM pre-trained model 1300 Machine Learning Devices 1301 CPU 1302 Storage medium 1303 Communication Interface 1311 Acquisition Department 1312 Learning Department 1312A Reasoning part 1312B Error calculation section 1312C Update Department 1313 Communication Control Unit 1321 Data Storage Unit 2500 Group Management Device

Claims

1. For each shot in which a molded product is formed by the injection molding machine, a first time-series data is acquired that shows the results detected by a detection device installed in the injection molding machine in chronological order. When the first time-series data is input to a machine learning model that has been machine-learned to estimate the time-series data of the results detected by the detection device for each shot in the injection molding machine, the machine learning model receives output data inferred as the time-series data of the results detected by the detection device during the time period corresponding to the first time-series data. A control unit that performs control to detect anomalies based on the difference between the first time-series data and the output data. A control device for an injection molding machine, equipped with the following features.

2. The machine learning model has an input layer, one or more intermediate layers connected to the input layer, and an output layer connected to the intermediate layers, and is formed as an encoder-decoder model. The control device for an injection molding machine according to claim 1.

3. The input to the input layer and the output to the output layer of the aforementioned machine learning model have a fixed data size. The control device for an injection molding machine according to claim 2.

4. The control unit, when the first time series data input as a result of detection from the detection device is larger than the fixed-length data size, deletes data from the first time series data to make it the fixed-length data size, and inputs the deleted first time series data into the machine learning model. The control device for an injection molding machine according to claim 3.

5. For each shot in which a molded product is formed by the injection molding machine, a first time-series data, which shows the results detected by a detection device installed in the injection molding machine in a time-series manner, is received from the injection molding machine. When the first time-series data is input to a machine learning model that has been machine-learned to estimate the time-series data of the results detected by the detection device for each shot in the injection molding machine, the machine learning model receives output data inferred as the time-series data of the results detected by the detection device during the time period corresponding to the first time-series data. A control unit that performs control to detect anomalies based on the difference between the first time-series data and the output data. A control device for an injection molding machine, equipped with the following features.

6. In an injection molding machine, a machine learning model is used to estimate the time series of results detected by a detection device for each shot. For each shot in which a molded product is formed by the injection molding machine, a first time series data representing the results detected by a detection device installed in the injection molding machine is input. A display unit then displays information based on the difference between the output data inferred from the machine learning model as the time series data of the results detected by the detection device during the time period corresponding to the first time series data, and the first time series data. A display device equipped with the following features.

7. Detection device and For each shot in which a molded product is formed by the device, a first time-series data is acquired that shows the results detected by the detection device in chronological order. When the first time-series data is input to a machine learning model that has been machine-learned to estimate the time-series data of the results detected by the detection device for each shot in the aforementioned device, the machine learning model receives output data inferred as the time-series data of the results detected by the detection device during the time period corresponding to the first time-series data. A control unit that performs control to detect anomalies based on the difference between the first time-series data and the output data, An injection molding machine equipped with [a specific feature / feature].

8. Detection device and A machine learning model, which has been trained to estimate the time series of results detected by a detection device for each shot in the device, receives first time series data, which shows the results detected by the detection device in time series, for each shot in which a molded product is formed by the device. A display unit displays information based on the difference between the output data inferred from the machine learning model as the time series data of the results detected by the detection device during the time period corresponding to the first time series data, and the first time series data. An injection molding machine equipped with [a specific feature / feature].

9. A machine learning device that enables machine learning of neural networks, A learning unit that, when inputting first time-series data showing the results detected by a detection device installed in an injection molding machine during one shot in which a molded product is formed by the injection molding machine, outputs a second time-series data identical to the first time-series data, uses the first time-series data to train a neural network. A machine learning device equipped with the following features.

Images