Lamination molding abnormality detection method, lamination molding abnormality device and program

Abstract

To provide a lamination molding abnormality detection method, a lamination molding abnormality detection device and a program capable of accurately and quickly detecting abnormality by catching a symptom of abnormality even if the determination of an abnormality occurrence is difficult.SOLUTION: A lamination molding abnormality detection method for repeatedly laminating a weld bead formed by melting a filler material by a heat input from a heat source acquires time series information related to an output of the heat source required in laminating the weld bead, acquires first abnormality degree information obtained by quantizing an abnormality degree of lamination molding on the basis of a first learning model that previously learns a relation between an output and the normality or abnormality of the lamination molding, extracts low abnormality degree data to be a threshold or below from the first abnormality degree information, acquires second abnormality degree information obtained by quantizing an abnormality degree of the time series information on the basis of transition in the extracted low abnormality degree data, and determines the abnormality of the lamination molding on the basis of the first abnormality degree information and the second abnormality degree information.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to an abnormality detection method for additive manufacturing, and an abnormality detection device and program for additive manufacturing. [Background technology]

[0002] As a method for determining the quality of a weld, a method for determining the quality of a weld based on a characteristic shape (signature) occurring in the waveform of a welding parameter such as a welding voltage or a welding current is disclosed in Patent Document 1. In this method, a defect analysis routine based on the signature shape can be used to determine whether the weld is acceptable or defective. In the defect analysis routine based on the signature shape, a reference welding signature for comparison with the captured welding signature is provided by a computing device. Then, the shape of the welding signature is compared with the reference welding signature shape, and a welding signature shape difference between the two is identified. This welding signature shape difference is an evaluation value that represents the degree of difference between the reference shape and the captured welding signature shape. Therefore, based on the identified welding signature shape difference, it is possible to determine whether the weld is acceptable or defective.

[0003] Furthermore, Patent Document 2 describes a method for determining welding abnormalities with high accuracy by utilizing a trained model that has learned parameters during normal welding. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2019-162666 A [Patent Document 2] JP 2019-188435 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, in additive manufacturing, weld beads may be formed along various types of modeling paths. For example, a modeling path of overhang stacking without a support structure, a modeling path that closes two overhanging parts, a modeling path that performs butt welding without a backing material, and even the positions of the start and end of the modeling path may each have a specific defect. For example, when forming a weld bead along a modeled overhanging part, the actual welding target position may deviate from the design position lined up with the existing weld bead, and the weld metal may adhere to an unplanned position, resulting in a defective shape. The same applies when forming beads in other cases, and defects such as arc interruption may occur. When such a defect occurs, it is required to immediately stop the modeling process and correct the defect, but there are cases where it is not possible to determine that an abnormality has occurred based on the difference in shape and the pre-trained model, so there is a demand for more accurate and rapid abnormality detection.

[0006] Therefore, an object of the present invention is to provide an anomaly detection method for additive manufacturing, an anomaly detection device, and a program for additive manufacturing that can capture signs of anomaly and detect anomalies accurately and quickly even when it is difficult to determine whether an anomaly has occurred. [Means for solving the problem]

[0007] The present invention comprises the following configurations. (1) A method for detecting an anomaly in additive manufacturing (AM) in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, comprising the steps of: Acquire time series information regarding the output of the heat source required to build up the weld bead; determining first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and based on the time-series information; extracting abnormality degree data that is equal to or smaller than a threshold value from the first abnormality degree information; determining second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; Anomaly detection method for additive manufacturing. (2) A method for detecting an anomaly in additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, comprising: Acquire time series information regarding the output of the heat source required to build up the weld bead; determining first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and based on the time-series information; extracting anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information obtained based on a model different from the first learning model and the time-series information; determining second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; Anomaly detection method for additive manufacturing. (3) An abnormality detection device for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, a heat source output information acquisition unit that acquires time-series information regarding the output of the heat source required when laminating the weld bead; a first abnormality degree calculation unit that calculates first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process, based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and on the time-series information; and an abnormality degree data extraction unit that extracts abnormality degree data in which the first abnormality degree information is equal to or smaller than a threshold; a second abnormality degree calculation unit that calculates second abnormality degree information by quantifying the degree of abnormality of the time-series information based on a transition in the extracted abnormality degree data; an abnormality determination unit that determines an abnormality in the additive manufacturing based on the first abnormality degree information and the second abnormality degree information; An additive manufacturing anomaly detection device comprising: (4) An abnormality detection device for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, a heat source output information acquisition unit that acquires time-series information regarding the output of the heat source required when laminating the weld bead; a first abnormality degree calculation unit that calculates first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process, based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and on the time-series information; and an anomaly degree data extraction unit that extracts anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information calculated based on a model different from the first learning model and the time-series information; a second abnormality degree calculation unit that calculates second abnormality degree information by quantifying the degree of abnormality of the time-series information based on a transition in the extracted abnormality degree data; an abnormality determination unit that determines an abnormality in the additive manufacturing based on the first abnormality degree information and the second abnormality degree information; An additive manufacturing anomaly detection device comprising: (5) A program for implementing an anomaly detection procedure for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, the program comprising: On the computer, acquiring time-series information regarding the output of the heat source required to build up the weld bead; determining first anomaly degree information that quantifies a degree of anomaly in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and the time-series information; A step of extracting abnormality degree data in which the first abnormality degree information is equal to or smaller than a threshold value; A step of calculating second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; A program to implement the above. (6) A program for implementing an anomaly detection procedure for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, the program comprising: On the computer, acquiring time-series information regarding the output of the heat source required to build up the weld bead; determining first anomaly degree information that quantifies a degree of anomaly in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and the time-series information; extracting anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information obtained based on a model different from the first learning model and the time-series information; A step of calculating second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; A program to implement the above. Effect of the Invention

[0008] According to the present invention, even when it is difficult to determine the occurrence of an abnormality, the symptom of the abnormality can be captured and the abnormality can be accurately detected. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a flowchart showing the procedure of the learning phase in the method for detecting anomalies in additive manufacturing. [Diagram 2] FIG. 2 is a flowchart showing the steps of the anomaly detection and monitoring phase in the anomaly detection method for additive manufacturing. [Diagram 3] FIG. 3 is a schematic diagram of the additive manufacturing system. [Figure 4] FIG. 4 is a schematic functional block diagram of the anomaly detection device. [Diagram 5] FIG. 5 is an explanatory diagram that illustrates the occurrence of burn-through. [Figure 6] Figure 6 is a photograph showing the state when burn-through occurs during molding. [Figure 7] FIG. 7 is an explanatory diagram showing waveforms of the welding voltage and the abnormality degree during molding before and after nozzle cleaning. [Figure 8] FIG. 8 is an explanatory diagram showing a procedure for detecting an anomaly using time-series data. [Figure 9] FIG. 9 is a photograph showing an example of a cross-sectional macro observation of a weld bead in a steady state when welding was performed under the CMT mode and pulse mode conditions. [Figure 10] FIG. 10 is a photograph showing an example of a cross-sectional macro observation of a weld bead in a steady state when welding was performed under the conditions of the CMT mode and the pulse mode. [Figure 11] FIG. 11 is a photograph showing an example of a cross-sectional macro observation of a weld bead in a steady state when welding was performed under the CMT mode and pulse mode conditions. [Figure 12] FIG. 12 is an explanatory diagram showing the comparison results of the voltage waveforms and the abnormality degrees in the pulse mode in Test Examples 1 to 12. [Figure 13] FIG. 13 is an explanatory diagram showing the comparison results of the current waveforms and the abnormality degrees in the pulse mode in Test Examples 1 to 12. [Figure 14] FIG. 14 is a graph showing the relationship between the mixture ratio of the shielding gas and the degree of abnormality of the welding current waveform. [Figure 15] FIG. 15 is a graph showing the relationship between the mixture ratio of the shielding gas and the degree of abnormality of the welding current waveform. [Figure 16] FIG. 16 is a graph showing the relationship between the mixture ratio of the shielding gas and the degree of abnormality of the welding voltage waveform. [Figure 17] FIG. 17 is a graph showing the relationship between the mixture ratio of the shielding gas and the degree of abnormality of the welding voltage waveform. [Figure 18] FIG. 18 is an explanatory diagram showing the change in the degree of abnormality depending on the presence or absence of the burn-through portion. [Figure 19] FIG. 19 is an explanatory diagram showing a change in the degree of abnormality in an adjacent pass to a modeling pass in which a burn-through portion occurs. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. The method for detecting anomalies in additive manufacturing according to this embodiment accurately and quickly detects anomalies that occur when forming a weld bead in additive manufacturing, in which a filler material is melted by heat input from a heat source to form a weld bead and the weld bead is repeatedly stacked to form a molded object. The anomalies referred to here are assumed to be phenomena or problems that interfere with the operation of the molding device or the continuation of the molding process. The procedure for this anomaly detection generally includes a learning phase and an anomaly detection / monitoring phase.

[0011] 1 is a flowchart showing the procedure of the learning phase in the method for detecting anomalies in additive manufacturing. In the learning phase, a test is performed to stack weld beads under predetermined test conditions such as bead-on-plate welding (S11), and time-series information such as the welding current and welding voltage during the test stacking is obtained (S12). A learning model (first learning model) is generated from this time-series information and the state (normal or abnormal) of the formed weld bead (S13).

[0012] FIG. 2 is a flowchart showing the procedure of the anomaly detection and monitoring phase in the anomaly detection method for additive manufacturing. In the anomaly detection and monitoring phase, time series information on the output of the heat source obtained during actual additive manufacturing is acquired (S21), and the acquired time series information is input to the first learning model prepared in the learning phase described above to obtain a quantitative degree of anomaly (first anomaly information) (S22). From the first anomaly information, it is possible to detect anomaly types that are easy to determine as having occurred from a pre-trained model. The degree of anomaly may be obtained from a model other than the first learning model and the time series information.

[0013] Furthermore, from the above-mentioned anomaly degree information, anomaly degree data below a threshold value is extracted (S23), and second anomaly degree information is obtained from the transition of this anomaly degree data (S24). From the second anomaly degree information, it is possible to detect signs of a malfunction that cannot be determined as an anomaly from the pre-trained model. The first anomaly degree information and the second anomaly degree information are combined to determine an anomaly in the additive manufacturing (S25).

[0014] <Details of the method for detecting anomalies in additive manufacturing> The steps of performing the additive manufacturing anomaly detection method using the first and second anomaly degrees, and the configuration of an additive manufacturing anomaly detection device (hereinafter referred to as an anomaly detection device) for detecting anomalies using the additive manufacturing anomaly detection method will be described below. First, an example of an additive manufacturing system that manufactures a molded object by stacking weld beads will be described with reference to FIG.

[0015] (Additive Manufacturing System) 3 is a schematic diagram of an additive manufacturing system 100. The additive manufacturing system 100 shown here forms a weld bead using an arc as a heat source. The additive manufacturing system 100 includes an additive manufacturing control device 11, a manipulator 13, a filler material supply device 15, a manipulator control device 17, and a heat source control device 19. The additive manufacturing control device 11 controls the above-mentioned components in an integrated manner.

[0016] The manipulator control device 17 controls the manipulator 13 and the heat source control device 19. A controller (not shown) is connected to the manipulator control device 17, and any operation of the manipulator control device 17 can be instructed by an operator via the controller.

[0017] The manipulator 13 is, for example, an articulated robot, and a torch 21 provided on the tip shaft is supported so that a filler material M can be continuously supplied. The torch 21 holds the filler material M in a state where it protrudes from the tip. The position and posture of the torch 21 can be set arbitrarily in three dimensions within the range of the degrees of freedom of the robot arm constituting the manipulator 13. The manipulator 13 is preferably one having six or more degrees of freedom, and is preferably one that can arbitrarily change the axial direction of the heat source at the tip. The manipulator 13 may be in various forms, such as a four- or more-axis articulated robot as shown in FIG. 1, or a robot equipped with angle adjustment mechanisms on two or more orthogonal axes.

[0018] The torch 21 has a shield nozzle (not shown), and a shielding gas is supplied from the shield nozzle. The shielding gas blocks the atmosphere and prevents oxidation and nitridation of the molten metal during welding, thereby suppressing poor welding. The arc welding method used in this configuration may be either a consumable electrode type such as a shielded metal arc welding or a carbon dioxide gas arc welding, or a non-consumable electrode type such as a TIG (Tungsten Inert Gas) welding or a plasma arc welding, and is appropriately selected depending on the object to be molded. Here, gas metal arc welding will be taken as an example for explanation. In the case of a consumable electrode type, a contact tip is disposed inside the shield nozzle, and a filler material M to which a current is supplied is held by the contact tip. The torch 21 generates an arc from the tip of the filler material M in a shielding gas atmosphere while holding the filler material M. Although not shown, a sensor capable of measuring the flow rate of a shielding gas such as Ar or CO2 gas may be provided, and a configuration may be used in which actual log data of the shielding gas flow rate is generated.

[0019] The filler metal supplying device 15 supplies the filler metal M toward the torch 21. The filler metal supplying device 15 includes a reel 15a on which the filler metal M is wound, and a pay-out mechanism 15b that pays out the filler metal M from the reel 15a. The filler metal M is fed to the torch 21 while being sent in the forward or reverse direction by the pay-out mechanism 15b as necessary. The pay-out mechanism 15b is not limited to a push type that is disposed on the filler metal supplying device 15 side and pushes out the filler metal M, but may be a pull type or a push-pull type that is disposed on a robot arm or the like.

[0020] The heat source control device 19 is a welding power source that supplies power required for welding by the manipulator 13. The heat source control device 19 adjusts the welding current and welding voltage supplied when forming a bead by melting and solidifying the filler metal M. Also, the filler metal supply speed of the filler metal supply device 15 is adjusted in conjunction with the welding conditions such as the welding current and welding voltage set by the heat source control device 19. Also, a configuration may be adopted in which sensors for the welding current, welding voltage, etc. are provided to generate performance log data of the actual welding current and welding voltage.

[0021] The heat source for melting the filler metal M is not limited to the above-mentioned arc. For example, other types of heat sources may be used, such as a heating method using both an arc and a laser, a heating method using plasma, or a heating method using an electron beam or a laser. When heating with an electron beam or a laser, the amount of heat can be controlled more precisely, and the state of the bead to be formed can be more appropriately maintained, which contributes to further improving the quality of the laminated structure.

[0022] The additive manufacturing system 100 having the above-mentioned configuration operates according to a modeling program created based on a modeling plan for the object 23. The modeling program is composed of a large number of command codes, and is created based on an appropriate algorithm according to various conditions such as the shape, material, and heat input of the object. When the torch 21 is moved and the supplied filler material M is melted and solidified according to this modeling program, a linear weld bead B, which is a molten solidified body of the filler material M, is formed on the base material 25. That is, the manipulator control device 17 drives the manipulator 13 and the heat source control device 19 based on a predetermined program provided from the modeling control device 11. The manipulator 13 moves the torch 21 while melting the filler material M with an arc, according to a command from the manipulator control device 17, to form the weld bead B. In this way, the weld beads B are sequentially formed and stacked to obtain a model 23 having a desired shape.

[0023] The molding control device 11 is configured by hardware using an information processing device such as a PC (Personal Computer).

[0024] When additive manufacturing of the object 23 is performed by the additive manufacturing system 100, the procedure is generally as follows. First, the target shape to be manufactured is sliced ​​at the height of the weld bead to generate a slice model of multiple layers. A torch movement path (modeling path) for forming beads with the width of the weld bead for each of the shapes of the generated slice models is obtained to generate a bead model simulating the weld bead. A weld bead B is formed along the bead model thus obtained, thereby forming a bead layer having the shape of the slice model. In this manner, a plurality of bead layers are repeatedly stacked to manufacture an object of the target shape. The cross-sectional shape of the weld bead B (bead width, bead height, etc.) is adjusted according to welding conditions such as the feed speed of the filler metal M, the welding voltage, and the welding current.

[0025] The above-mentioned modeling program includes information on the modeling path and welding conditions for forming the above-mentioned weld bead B. The modeling control device 11 outputs a modeling program corresponding to the shape of the object 23 to be modeled to the manipulator control device 17, and causes the manipulator control device 17 to model the desired shape of the object 23.

[0026] (Anomaly detection device) An anomaly detection device 200 is connected to the additive manufacturing system 100. The anomaly detection device 200 is configured by hardware using an information processing device such as a PC (Personal Computer) like the manufacturing control device 11. Each function of the anomaly detection device 200 is realized by a control unit (not shown) reading out a program having a specific function stored in a storage device (not shown) and executing the program. Examples of the storage device include a memory such as a random access memory (RAM) which is a volatile storage area, a read only memory (ROM) which is a non-volatile storage area, and a storage such as a hard disk drive (HDD) or a solid state drive (SSD). Examples of the control unit include a processor such as a central processing unit (CPU) or a micro processor unit (MPU), or a dedicated circuit.

[0027] The anomaly detection device 200 is not limited to being directly connected to the additive manufacturing system 100, and may be configured as another computer remotely connected via a network or the like.

[0028] 4 is a schematic functional block diagram of the anomaly detection device. The anomaly detection device 200 includes a heat source output information acquisition unit 31, a first anomaly degree calculation unit 33, an anomaly degree data extraction unit 35, a second anomaly degree calculation unit 37, and an anomaly determination unit 39. The anomaly detection device 200 may also include a partial data extraction unit 61 and a second learning model generation unit 63, the details of which will be described later.

[0029] Heat source output information acquisition unit 31 acquires time-series information 41 relating to the output of the heat source required to build up a weld bead. The heat source in this configuration is an arc adjusted by heat source control device 19, and the time-series information acquired includes information such as set values ​​of the welding current, welding voltage, etc. supplied by heat source control device 19 when forming a bead, or recorded log data (actual log data). When a laser is used as the heat source, time-series information such as the transition of the reflected light intensity of the laser may also be included.

[0030] The first anomaly degree calculation unit 33 calculates first anomaly degree information 45 that quantifies the degree of anomaly in the additive manufacturing based on a first learning model 43 that has been generated by previously learning the relationship between the output of the heat source and whether the additive manufacturing is normal or abnormal, and the acquired time-series information 41. In other words, the time-series information 41 is input to a first learning model 43 prepared in advance, and first anomaly degree information 45 that indicates the degree of anomaly in the additive manufacturing is calculated from the characteristics of the time-series information 41.

[0031] When generating the first abnormality degree information 45, noise removal, selection of a data interval to be learned, labeling such as normal / abnormal, etc. may be arbitrarily performed. When the first learning model 43 is a classification model such as abnormal (1) and normal (0), a value obtained by aggregating the number of abnormality judgments occurring in a certain time unit may be set as the abnormality degree information. For example, the K Nearest Neighbor method or other known machine learning methods may be used to generate the first learning model 43. In addition, a learning method that outputs a quantitative value may be applied other than the above. For example, for time series data x(t), a learning model (generally called an autoregressive model) may be configured to learn the relationship when the input is x(t-3), x(t-2), and x(t-1) and the output is x(t) to predict the output after a certain time has passed from the history of the time series data.

[0032] The anomaly data extraction unit 35 extracts anomaly data 47 that is equal to or smaller than a predetermined threshold from the first anomaly information 45. Alternatively, the anomaly data 47 may be extracted from a model other than the first learning model and time-series information. This extraction may be performed by converting the time transition of the anomaly determination result into a filter and applying the filter to the time-series information. The anomaly data 47 extracted here may include, in addition to data when there is no problem in terms of the modeling, signs of defects that cannot be determined to be abnormal from the pre-learned model. Such signs are identified by the second anomaly calculation unit 37.

[0033] The second anomaly degree calculation unit 37 calculates second anomaly degree information 49 by quantifying the degree of anomaly of the time-series information 41 based on the transition in the extracted anomaly degree data 47. From this second anomaly degree information 49, a symptom of a malfunction that cannot be determined as an anomaly occurrence from the first learning model 43 that has been trained in advance is detected.

[0034] For example, an average behavior may be identified from a portion of the extracted anomaly data 47, and behavior that deviates from the average behavior may be estimated as abnormal. The average behavior here may be a waveform of a specific period extracted from the extracted data, or a waveform whose shape matches the waveforms of the welding current and welding voltage observed in the past modeling history. Alternatively, the average behavior may simply be a moving average value of the extracted data. As a method for determining an abnormality, the above-mentioned average behavior may be compared with the value of the data in the section to be detected, and the second anomaly information 49 may be calculated from the magnitude of the difference.

[0035] Also, a time interval showing the average behavior may be used as training data to generate a second learning model 51 by a learning method such as the K-nearest neighbor method, and the extracted anomaly degree data 47 may be input to this second learning model 51 to determine second anomaly degree information 49, which is the degree of deviation from the average behavior. Alternatively, a prediction model (the above-mentioned autoregressive model) may be trained to predict the welding current and welding voltage a predetermined time after the average behavior is shown, and the difference between the predicted value by the prediction model and the actual measured value, or the amount of deviation such as the squared value of the difference, may be determined as the second anomaly degree information 49.

[0036] The anomaly determination unit 39 comprehensively determines anomalies in the additive manufacturing process based on the first anomaly degree information 45 and the second anomaly degree information 49. Since the first anomaly degree information 45 is compared with anomalies previously learned, if the first anomaly degree information 45 is greater than a predetermined threshold, it may be determined that an anomaly exists regardless of the second anomaly degree information. In this case, it is relatively easy to identify the cause of the anomaly. On the other hand, if the first anomaly degree information 45 is less than the threshold and the second anomaly degree information 49 exceeds the predetermined threshold, it is suspected that an unknown anomaly has occurred. Therefore, it may be determined that an anomaly has occurred under limited conditions, such as when the second anomaly degree information 49 continues to exceed the threshold for a predetermined time period, or when the control conditions are corrected or changed. If neither the first anomaly degree information 45 nor the second anomaly degree information 49 exceeds the threshold, it may be determined that the additive manufacturing process is normal.

[0037] <Major types of defects that occur during additive manufacturing> (1) Insufficient shielding gas flow rate, inappropriate shielding gas mixture ratio It is assumed that defects related to shielding gas occur due to hardware problems such as gas supply failure, nozzle blockage, gas cylinder empty, valve not opening, etc., as well as human error. It is generally known that defects such as instability of the shapes of the arc, molten pool, and bead, increase in spatter, and occurrence of blowhole defects occur when the shielding gas flow rate is insufficient or the mixture ratio of Ar gas and CO2 gas is inappropriate. Since abnormalities in the shielding gas flow rate and mixture ratio can be relatively easily reproduced, by conducting tests to investigate the effects of these flow rates and mixture ratios and subjecting the results to machine learning in advance, abnormalities caused by shielding gas can be easily identified by the first anomaly degree.

[0038] (2) Meltdown FIG. 5 is an explanatory diagram showing a schematic diagram of the occurrence of burn-through. FIG. 6 is a photograph showing the occurrence of burn-through during the manufacturing process. As shown in FIG. 5, when forming a weld bead with the tip of an overhanging molded object 23 as the target position P, if the torch 21 deviates from the target position P in the manufacturing plan to the position of the broken line, the weld metal melts through from the target position P and adheres to an unplanned location Pa. If such burn-through (also called "watermark") occurs during welding or arc-on, the molded object 23 will have a defective shape, so a molding technique that avoids such an abnormality as much as possible is basically required. In addition, if an abnormality occurs due to an unexpected malfunction, it is desirable to immediately and automatically stop the manufacturing process, and ideally, it is desirable to automatically stop the manufacturing process before an abnormality occurs. There are a great many patterns regarding the cause of burn-through depending on the layer shape, the target position setting of the torch, the attitude of the torch, and the like, and a reproduction test is also relatively difficult. Therefore, it is preferable to estimate the occurrence of burn-through comprehensively from the first abnormality degree and the second abnormality degree.

[0039] (3) Torch nozzle sputter deposition When spatter generated during additive manufacturing accumulates in the torch nozzle, it gradually blocks the shielding gas outlet, which can have a negative effect on the generation of the arc. If an abnormality occurs due to spatter attached to the torch nozzle, it is necessary to interrupt the manufacturing process and clean or replace the torch nozzle. This abnormality is relatively easy to reproduce. Therefore, by acquiring the waveforms of the welding current and welding voltage when the amount of spatter accumulated on the torch nozzle changes and providing them as learning data for the machine learning model, an abnormality caused by slag accumulation on the torch nozzle can be identified by the first anomaly degree.

[0040] Figure 7 is an explanatory diagram showing the waveforms of the welding voltage and abnormality during molding before and after nozzle cleaning. Before nozzle cleaning, large fluctuations in repeated voltage increases occur in the ST portion of the welding voltage waveform near the end of the pass, and the abnormality is high in this portion. On the other hand, after nozzle cleaning, the increasing fluctuations in the welding voltage are reduced, and the increase in the abnormality is also suppressed. In this way, the welding voltage waveform changes due to spatter adhering to the torch nozzle, so abnormalities due to slag volume can be identified from the welding voltage waveform.

[0041] <Method of detecting abnormal occurrence> The occurrence of abnormalities such as those described above is detected using time-series information such as the welding current waveform or welding voltage waveform obtained by measurement during molding. The abnormality detection here utilizes a machine learning method (K-nearest neighbor method) to calculate an index (degree of abnormality) that indicates the difference between normal and abnormal, thereby determining whether an abnormality has occurred.

[0042] The K-nearest neighbor method is a method for calculating the degree of anomaly from the ratio of abnormal data among k pieces of data near the data to be discriminated, and does not require any preconditions for data distribution, and can be applied even if normal data consists of multiple sets. In addition, it has the advantage that the degree of anomaly can be calculated easily and the required calculation cost is small.

[0043] 8 is an explanatory diagram showing the procedure for detecting anomalies from arbitrary time-series information. The data processing process using the K-nearest neighbor method includes the following steps (1) to (3). (1) As shown in Fig. 8, a waveform of a specific section among the transition waveforms of data 36, ​​which is time-series information, is set as training data 55. In addition, waveforms of other sections corresponding to the waveforms of training data 55 are sequentially extracted from the transition waveform of data 36 as detection data 57. The section (win(1)) of training data 55 and the section (win(2)) of detection data 57 are set to have the same width.

[0044] The training data 55 and the detection data 57 are extracted by a partial data extraction unit 61 shown in Fig. 4. The data 36 input to the partial data extraction unit 61 may be the time-series information 41 from the heat-source output information acquisition unit 31, the first anomaly degree information 45 from the first anomaly degree calculation unit 33, or the anomaly degree data 47 from the anomaly degree data extraction unit 35. Then, the partial data extraction unit 61 outputs the extracted training data 55 and detection data 57 to the second learning model generation unit 63.

[0045] (2) The second learning model generation unit 63 generates the second learning model 51 from the extracted detection data 57 and training data 55 using a learning method such as the K-nearest neighbor method. Then, the second anomaly degree calculation unit 37 inputs the anomaly degree data 47 to the second learning model 51 to obtain the second anomaly degree information 49. Alternatively, the second anomaly degree information 49 may be obtained from the magnitude of a feature amount such as a difference value between the detection data 57 and the training data 55. In this specification, the process of obtaining the degree of anomaly from the magnitude of such a feature amount is also referred to as a process using a "model". Specifically, the squared distance between the training data 55 and each detection data 57 may be obtained, and the minimum value a of the squared distances may be used as the second anomaly degree information 49. In this case, the squared distance dn(x) is obtained by equation (1), and the minimum value a is obtained by equation (2).

[0046]

number

[0047] (3) The abnormality determination unit 39 determines that an abnormality has occurred when the second abnormality degree information 49 calculated by the second abnormality degree calculation unit 37 is equal to or greater than a predetermined threshold value, even if the first abnormality degree information 45 is at a normal level.

[0048] In the anomaly detection device 200 of this configuration, the second anomaly information 49 is obtained in addition to the first anomaly information 45, and the first anomaly information 45 and the second anomaly information 49 are used to determine the occurrence of an anomaly during modeling. By performing this determination in real time, in-line monitoring is possible. Therefore, when an anomaly occurs, modeling can be stopped immediately. In addition, when the performance log data is generated, the cause of the anomaly can be estimated from the performance log data, and the anomaly can be verified after modeling. Then, by calculating the first anomaly information 45 and the second anomaly information 49 using algorithms such as machine learning and statistical models, the accuracy of the anomaly calculation can be improved, and the effect of suppressing overdetection can be expected. In addition, by analyzing the occurrence of an anomaly and its cause, it becomes easy to automate the taking of measures to avoid an anomaly to minimize modeling defects after an anomaly is detected, and highly efficient and high-quality modeling can be realized. EXAMPLES

[0049] <Changes in shielding gas supply conditions> Abnormalities in the supply state of the shielding gas include insufficient total gas flow rate or an inappropriate mixture ratio of Ar gas and CO2 gas. Therefore, bead-on-plate (BOP) tests were carried out under typical welding conditions with gas supply conditions simulating these conditions, and the effects of differences in gas supply conditions on the behavior and bead properties during welding, as well as correlations with various measurement data, were verified. Test examples 1 to 12, in which the gas supply conditions were changed, are summarized in Table 1. The standard extension length of the welding wire was 15 mm (Test examples 1 to 10).

[0050] [Table 1]

[0051] Here, as typical welding conditions, the cold metal transfer (CMT: registered trademark) mode and the pulse mode were carried out under the following conditions. (CMT mode) Wire feed speed: 1.5mpm Welding speed: 15cpm No weaving (Pulse mode) Wire feed speed: 9.5mpm Welding speed: 75cpm No weaving

[0052] 9 to 11 are photographs showing examples of cross-sectional macro observations of weld beads in the steady state when welding was performed under CMT mode and pulse mode conditions. Fig. 9 is a cross-sectional photograph of Test Example 1, which is the reference condition, Fig. 10 is a cross-sectional photograph of Test Example 2, in which the shielding gas was 100% CO2 gas, and Fig. 11 is a cross-sectional photograph of Test Example 10, in which the shielding gas was 100% Ar gas.

[0053] 9 to 11, it can be confirmed that differences in the bead shape and penetration state occur depending on the shielding gas supply conditions. Moreover, no blowholes were observed under any of the conditions of Test Examples 1 to 12.

[0054] Fig. 12 is an explanatory diagram showing the results of comparing the voltage waveform and the degree of anomaly in pulse mode in Test Examples 1 to 12. Fig. 13 is an explanatory diagram showing the results of comparing the current waveform and the degree of anomaly in pulse mode in Test Examples 1 to 12. The standardized degree of anomaly in the rightmost columns of Figs. 12 and 13 is a value obtained by standardizing the degree of anomaly calculated for the waveform under each condition by the K-nearest neighbor method described above, with the calculation result for Test Example 1, which is the reference condition. In other words, the larger the standardized degree of anomaly, the greater the deviation from the waveform obtained under the reference condition.

[0055] In test examples 2, 3, 4, 6, 8, and 10, the mixture ratio of the shielding gas was changed from that of test example 1, which was the reference condition. In that case, the waveform of the welding current was clearly different from that of the reference condition, and a significant increase in the fluctuation of the waveform of the welding voltage was confirmed in test examples 2, 3, and 10, where the difference in the mixture ratio from the reference condition was large. In each of the above test examples, the abnormality value increased. In test examples 5, 7, and 9, the mixture ratio of the shielding gas was kept the same as that of test example 1, which was the reference condition, while the total flow rate of the shielding gas was changed. In that case, the waveform of the welding current and welding voltage did not differ as much as described above, and the abnormality value was the same or only slightly increased. In test examples 11 and 12, the mixture ratio of the shielding gas was kept the same as that of test example 1, which was the reference condition, while the total flow rate of the shielding gas was changed. In addition, the protrusion amount of the welding wire was changed. In that case, some disturbance occurred in the waveform of the welding voltage, but the current waveform was the same as that of the reference condition, and the increase in the abnormality value was also slight.

[0056] Figures 14 and 15 are graphs showing the relationship between the shielding gas mixture ratio and the degree of abnormality in the welding current waveform. Figure 14 shows the test results in CMT mode, and Figure 15 shows the test results in pulse mode. Figures 16 and 17 are graphs showing the relationship between the shielding gas mixture ratio and the degree of abnormality in the welding voltage waveform. Figure 16 shows the test results in CMT mode, and Figure 17 shows the test results in pulse mode.

[0057] As shown in Figures 14 to 17, when the shielding gas mixing ratio is the same as the shielding gas mixing ratio under the reference conditions (CO2 / Ar = 0.25), the increase in the abnormality value is small in both the CMT mode and the pulse mode, but the abnormality value increases as the shielding gas mixing ratio deviates from the mixing ratio under the reference conditions.

[0058] These results suggest that it is highly possible to detect abnormalities in the shielding gas supply by monitoring changes in the waveforms of the welding current and welding voltage during welding.

[0059] <Welding current waveform when burn-through occurs during molding> Next, we show a modeling pass in which base burn-through occurred due to insufficient base thickness during the part manufacturing process, and the changes in the welding current waveform and abnormality in the adjacent pass adjacent to the modeling pass in which the burn-through occurred. Here, we assume that burn-through occurred during crater treatment at the end of the weld bead, and the arc start point of the subsequent adjacent pass was located near the burn-through.

[0060] Fig. 18 is an explanatory diagram showing the change in the degree of abnormality depending on the presence or absence of the burn-through portion. Fig. 18 shows the results of analyzing the normal welding current waveform and the degree of abnormality using the K-nearest neighbor method when the normal welding current waveform is used as the first verification data, and the results of analyzing the welding current waveform when the burn-through portion occurs and the degree of abnormality using the K-nearest neighbor method when the normal welding current waveform is used as the first verification data, in a modeling path different from the training data.

[0061] In the first verification data, the degree of abnormality was slightly higher in the initial part of the welding current waveform where the waveform differed from the training data due to crater processing control, etc., but in the latter part of the welding current waveform, the difference with the training data was small and the degree of abnormality was small. In the second verification data, a period occurred where the welding current did not flow, and the degree of abnormality increased at that timing.

[0062] Fig. 19 is an explanatory diagram showing the change in the degree of abnormality in an adjacent pass to a printing pass in which a burn-through portion occurred. Fig. 19 shows the results of analyzing the normal welding current waveform before the burn-through occurred and the degree of abnormality using the K-nearest neighbor method when the normal welding current waveform before the burn-through occurred and the first verification data in a printing pass other than the training data, and the results of analyzing the welding current waveform when the burn-through portion occurred and the second verification data in a printing pass other than the training data.

[0063] In the first verification data in this case, the degree of abnormality was slightly high at the beginning of the welding current waveform, as in the case shown in Fig. 18, but the degree of abnormality was small in the latter half of the welding current waveform. In the second verification data, a period in which the welding current did not flow (the area surrounded by the dashed line in Fig. 19) occurred, and the degree of abnormality increased at that timing.

[0064] From the above results, it was confirmed that the degree of abnormality was high in the forming pass where burn-through occurred. It was also confirmed that if arc interruption (burn-through) occurs early in welding near the arc start of an adjacent pass adjacent to the area where burn-through occurred, this change can be detected by the degree of abnormality. In this way, it was shown that it is possible to detect multiple types of abnormalities, such as burn-through, in addition to incompatibility of the shielding gas supply condition, from time-series data such as welding voltage and welding current.

[0065] As described above, the present specification discloses the following: (1) A method for detecting an anomaly in additive manufacturing (AM) in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, comprising the steps of: Acquire time series information regarding the output of the heat source required to build up the weld bead; determining first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and based on the time-series information; extracting abnormality degree data that is equal to or smaller than a threshold value from the first abnormality degree information; determining second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; Anomaly detection method for additive manufacturing. According to this method for detecting anomalies in additive manufacturing, normal data is extracted using a pre-trained first learning model, and then the degree of anomaly is output from the progression of the normal data, making it possible to detect signs of abnormal phenomena that are difficult to detect using the pre-trained first model.

[0066] (2) A method for detecting an anomaly in additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, comprising: Acquire time series information regarding the output of the heat source required to build up the weld bead; determining first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and based on the time-series information; extracting anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information obtained based on a model different from the first learning model and the time-series information; determining second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; Anomaly detection method for additive manufacturing. According to this method for detecting anomalies in additive manufacturing, anomalies are detected using a pre-trained first learning model, and then anomalies are detected from the trends in anomaly data obtained using other models, making it possible to determine signs of abnormal phenomena that are difficult to determine using the pre-trained first learning model.

[0067] (3) setting a waveform of a specific section of the transition waveform of the anomaly degree data as training data, and sequentially extracting waveforms of other sections corresponding to the waveform of the training data from the transition waveform as detection data; generating a second learning model that learns features of the extracted detection data and the training data; The method for detecting an anomaly in additive manufacturing described in (1) or (2), further comprising inputting the anomaly data into the second learning model to obtain the second anomaly information. According to this method for detecting anomalies in additive manufacturing, a portion is sampled from the data or time series information extracted from the first learning model and its tendency is learned, so that when a unique tendency that is not similar to other sections appears, this is reflected in the second anomaly degree, and the occurrence of an anomaly is detected.

[0068] (4) A method for detecting an anomaly in additive manufacturing described in any one of (1) to (3), wherein the second learning model includes an autoregressive model that predicts the output after a certain period of time from the output history of the heat source based on information on the section used as the training data, and the second anomaly information is calculated based on the difference between the predicted output value and the actually observed output value. According to this method for detecting anomalies in additive manufacturing, the welding current and welding voltage during normal times fluctuate periodically, so if there are no abnormalities, future welding current and welding voltage can be easily predicted. Therefore, the deviation from the actual measured values ​​can be regarded as the degree of anomaly. In addition, it becomes easier to extract anomalies even if they occur over a relatively short period of time.

[0069] (5) A method for detecting an anomaly in additive manufacturing described in any one of (1) to (4), wherein the first learning model is a model that has learned whether the additive manufacturing is normal or abnormal due to fluctuations in the shielding gas mixing ratio. According to this method for detecting anomalies in additive manufacturing, anomalies that occur when the shielding gas mixture ratio is changed are abnormal phenomena that can be easily reproduced in simple tests such as bead-on-plate (BOP) tests, so by having the model learn in advance, it becomes easier to extract anomalies. In addition, when the first anomaly level exceeds the threshold, it is easy to suspect an anomaly in the shielding gas, making it easier to identify the cause of the anomaly.

[0070] (6) A method for detecting an anomaly in additive manufacturing described in any one of (1) to (5), which estimates the occurrence of a burn-through phenomenon of weld metal based on a pattern of the second anomaly degree output in association with the stacking of adjacent weld beads. According to this method for detecting anomalies in additive manufacturing, an abnormality involving burn-through of a weld bead affects not only the manufacturing pass of that weld bead but also other manufacturing passes, so the occurrence of the burn-through phenomenon can be inferred from the degree of abnormality between the manufacturing passes of adjacent weld beads.

[0071] (7) An abnormality detection device for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, a heat source output information acquisition unit that acquires time-series information regarding the output of the heat source required when laminating the weld bead; a first abnormality degree calculation unit that calculates first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process, based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and on the time-series information; and an abnormality degree data extraction unit that extracts abnormality degree data in which the first abnormality degree information is equal to or smaller than a threshold; a second abnormality degree calculation unit that calculates second abnormality degree information by quantifying the degree of abnormality of the time-series information based on a transition in the extracted abnormality degree data; an abnormality determination unit that determines an abnormality in the additive manufacturing based on the first abnormality degree information and the second abnormality degree information; An additive manufacturing anomaly detection device comprising: This additive manufacturing anomaly detection device extracts normal data using a pre-trained first learning model, and then outputs the degree of anomaly from the progression of the normal data, making it possible to detect signs of abnormal phenomena that are difficult to detect using the pre-trained first model.

[0072] (8) An abnormality detection device for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, a heat source output information acquisition unit that acquires time-series information regarding the output of the heat source required when laminating the weld bead; a first abnormality degree calculation unit that calculates first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process, based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and on the time-series information; and an anomaly degree data extraction unit that extracts anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information calculated based on a model different from the first learning model and the time-series information; a second abnormality degree calculation unit that calculates second abnormality degree information by quantifying the degree of abnormality of the time-series information based on a transition in the extracted abnormality degree data; an abnormality determination unit that determines an abnormality in the additive manufacturing based on the first abnormality degree information and the second abnormality degree information; An additive manufacturing anomaly detection device comprising: With this additive manufacturing anomaly detection device, anomalies are detected using a pre-trained first learning model, and then anomalies are detected from the trends in anomaly data obtained using other models, making it possible to detect signs of abnormal phenomena that are difficult to detect using the pre-trained first learning model.

[0073] (9) a partial data extraction unit that sets a waveform of a specific section of the transition waveform of the anomaly level data as training data, and sequentially extracts waveforms of other sections corresponding to the waveform of the training data from the transition waveform as detection data; a second learning model generation unit that generates a second learning model that has learned features of the extracted detection data and the training data; Equipped with The additive manufacturing anomaly detection device according to (7) or (8), wherein the second anomaly degree calculation unit inputs the anomaly degree data into the second learning model to obtain the second anomaly degree information. According to this additive manufacturing anomaly detection device, a portion of the obtained anomaly data is sampled and its trend is learned, so that when a unique trend that is not similar to other sections appears, this is reflected in the second anomaly level, and the occurrence of an anomaly is detected.

[0074] (10) A program for performing an anomaly detection procedure for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, the program comprising: On the computer, acquiring time-series information regarding the output of the heat source required to build up the weld bead; determining a first abnormality degree by quantifying a degree of abnormality of the additive manufacturing, based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing is normal or abnormal, and based on the time-series information; A step of extracting low abnormality degree data in which the first abnormality degree is equal to or less than a threshold value from the time-series information; calculating a second degree of anomaly by quantifying a degree of anomaly in the time-series information based on a transition in the extracted low degree of anomaly data; determining an abnormality in the additive manufacturing process based on the first abnormality degree and the second abnormality degree; A program to implement the above. According to this program, normal data is extracted using a pre-trained model, and then the degree of abnormality is output from the trends in that normal data, making it possible to detect signs of abnormal phenomena that are difficult to detect using a pre-trained model.

[0075] (11) A program for performing an anomaly detection procedure for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, the program comprising: On the computer, acquiring time-series information regarding the output of the heat source required to build up the weld bead; determining first anomaly degree information that quantifies a degree of anomaly in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and the time-series information; extracting anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information obtained based on a model different from the first learning model and the time-series information; A step of calculating second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; A program to implement the above. According to this program, anomalies are detected using a pre-trained first learning model, and then anomalies are detected from trends in anomaly data obtained using other models, making it possible to identify signs of abnormal phenomena that are difficult to detect using the pre-trained first learning model.

[0076] (12) The program according to (10) or (11), a step of setting a waveform of a specific section of the transition waveform of the anomaly level data as training data, and sequentially extracting waveforms of other sections corresponding to the waveform of the training data from the transition waveform as detection data; generating a second learning model that learns features of the extracted detection data and the training data; A step of inputting the anomaly degree data into the second learning model to obtain the second anomaly degree information; A program to further implement the above. According to this program, a portion of the anomaly data is sampled and its tendency is learned, so that when a peculiar tendency that is not similar to other sections appears, it is reflected in the second anomaly degree, and the occurrence of an anomaly is detected. [Explanation of symbols]

[0077] 11. Modeling control device 13 Manipulator 15 Filler metal supply device 15a Reel 15b Payout mechanism 17 Manipulator control device 19 Heat source control device 21 Torch 23 Sculptures 25 Base material 31 Heat source output information acquisition section 33 First abnormality calculation unit 35 Anomaly Data Extraction Unit 36 Data 37 Second abnormality calculation section 39 Abnormality determination section 41 Time Series Information 43 First Learning Model 45 First abnormality information 47 Anomaly Data 49 Second abnormality information 51 Second Learning Model 55 Training Data 57 Detection Data 100 Additive Manufacturing System 200 Anomaly Detection Device

Claims

1. A method for detecting an abnormality in additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, comprising: Acquire time series information regarding the output of the heat source required to build up the weld bead; determining first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and based on the time-series information; extracting abnormality degree data that is equal to or smaller than a threshold value from the first abnormality degree information; determining second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; Anomaly detection method for additive manufacturing.

2. A method for detecting an abnormality in additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, comprising: Acquire time series information regarding the output of the heat source required to build up the weld bead; determining first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and based on the time-series information; extracting anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information obtained based on a model different from the first learning model and the time-series information; determining second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; Anomaly detection method for additive manufacturing.

3. A waveform of a specific section of the transition waveform of the anomaly level data is set as training data, and waveforms of other sections corresponding to the waveform of the training data are sequentially extracted as detection data from the transition waveform; generating a second learning model that learns features of the extracted detection data and the training data; inputting the anomaly degree data into the second learning model to obtain the second anomaly degree information; The method for detecting an anomaly in additive manufacturing according to claim 1.

4. A waveform of a specific section of the transition waveform of the anomaly level data is set as training data, and waveforms of other sections corresponding to the waveform of the training data are sequentially extracted as detection data from the transition waveform; generating a second learning model that learns features of the extracted detection data and the training data; inputting the anomaly degree data into the second learning model to obtain the second anomaly degree information; The method for detecting an anomaly in additive manufacturing according to claim 2.

5. the second learning model includes an autoregressive model that predicts the output after a certain time has elapsed from the output history of the heat source based on information of the section used as the training data, and calculates the second anomaly degree information based on a difference between the predicted value of the output and an actually observed output value. The method for detecting an anomaly in additive manufacturing according to claim 3.

6. the second learning model includes an autoregressive model that predicts the output after a certain time has elapsed from the output history of the heat source based on information of the section used as the training data, and calculates the second anomaly degree information based on a difference between the predicted value of the output and an actually observed output value. The method for detecting an anomaly in additive manufacturing according to claim 4.

7. The first learning model is a model that learns normality or abnormality of the additive manufacturing due to fluctuations in the shielding gas mixing ratio, The method for detecting an abnormality in additive manufacturing according to any one of claims 1 to 6.

8. Inferring the occurrence of a burn-through phenomenon of a weld metal based on a pattern of the second abnormality information output in association with stacking of the adjacent weld beads. The method for detecting an abnormality in additive manufacturing according to any one of claims 1 to 6.

9. An abnormality detection device for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, a heat source output information acquisition unit that acquires time-series information regarding the output of the heat source required when laminating the weld bead; a first abnormality degree calculation unit that calculates first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process, based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and on the time-series information; and an abnormality degree data extraction unit that extracts abnormality degree data in which the first abnormality degree information is equal to or smaller than a threshold; a second abnormality degree calculation unit that calculates second abnormality degree information by quantifying the degree of abnormality of the time-series information based on a transition in the extracted abnormality degree data; an abnormality determination unit that determines an abnormality in the additive manufacturing based on the first abnormality degree information and the second abnormality degree information; An additive manufacturing anomaly detection device comprising:

10. An abnormality detection device for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, a heat source output information acquisition unit that acquires time-series information regarding the output of the heat source required when laminating the weld bead; a first abnormality degree calculation unit that calculates first abnormality degree information that quantifies a degree of abnormality in the additive manufacturing process, based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and on the time-series information; and an anomaly degree data extraction unit that extracts anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information calculated based on a model different from the first learning model and the time-series information; a second abnormality degree calculation unit that calculates second abnormality degree information by quantifying the degree of abnormality of the time-series information based on a transition in the extracted abnormality degree data; an abnormality determination unit that determines an abnormality in the additive manufacturing based on the first abnormality degree information and the second abnormality degree information; An additive manufacturing anomaly detection device comprising:

11. a partial data extraction unit that sets a waveform of a specific section of the transition waveform of the anomaly level data as training data, and sequentially extracts waveforms of other sections corresponding to the waveform of the training data from the transition waveform as detection data; a second learning model generation unit that generates a second learning model that has learned features of the extracted detection data and the training data; Equipped with the second abnormality degree calculation unit inputs the abnormality degree data into the second learning model to obtain the second abnormality degree information; The additive manufacturing abnormality detection device according to claim 9 or 10.

12. A program for performing an anomaly detection procedure for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, On the computer, acquiring time-series information regarding the output of the heat source required to build up the weld bead; determining first anomaly degree information that quantifies a degree of anomaly in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and the time-series information; A step of extracting abnormality degree data in which the first abnormality degree information is equal to or smaller than a threshold value; A step of calculating second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; A program to implement the above.

13. A program for performing an anomaly detection procedure for additive manufacturing in which a weld bead formed by melting a filler material with heat input from a heat source is repeatedly stacked, On the computer, acquiring time-series information regarding the output of the heat source required to build up the weld bead; determining first anomaly degree information that quantifies a degree of anomaly in the additive manufacturing process based on a first learning model that has previously learned a relationship between the output and whether the additive manufacturing process is normal or abnormal, and the time-series information; extracting anomaly degree data that is equal to or smaller than a threshold value from anomaly degree information obtained based on a model different from the first learning model and the time-series information; A step of calculating second anomaly degree information that quantifies the degree of anomaly of the time-series information based on a transition in the extracted anomaly degree data; determining an abnormality in the additive manufacturing process based on the first abnormality degree information and the second abnormality degree information; A program to implement the above.

14. The program according to claim 12 or 13, a step of setting a waveform of a specific section of the transition waveform of the anomaly level data as training data, and sequentially extracting waveforms of other sections corresponding to the waveform of the training data from the transition waveform as detection data; generating a second learning model that learns features of the extracted detection data and the training data; A step of inputting the anomaly degree data into the second learning model to obtain the second anomaly degree information; A program to further implement the above.

Images