Computer program, anomaly detection method, anomaly detection device, molding machine system, and learning model generation method

Abstract

To provide an abnormality detection device for providing a computer program which is capable of converting time series data of a physical quantity related to a movable part of an industrial machine into a time-series data image expressed by an image, and determining an abnormality of the industrial machine based on the time-series data image.SOLUTION: Provided is a computer program for causing a computer to execute processing of detecting an abnormality of an industrial machine having a movable section, the processing executed by the computer including: acquiring physical amount data, of a time series, outputted from a sensor that detects a physical amount relating to movement of the movable section; converting the acquired physical amount data of the time series to a time series data image expressed by an image; inputting the converted time series data image into a learning model that has been trained by using features of the time series data image relating to a normal movable section, and thereby calculating a feature amount of the time series data image; and determining presence or absence of an abnormality in the industrial machine on the basis of the calculated feature amount.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a computer program for detecting an abnormality in an industrial machine having a movable part, an abnormality detection method, an abnormality detection device, a molding machine system, and a learning model generation method.

Background Art

[0002] Patent Document 1 discloses a monitoring method for monitoring the vibration of each movable part of an injection molding machine with an acceleration sensor and detecting the state of the molding process and the occurrence of an abnormality in the movable part.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, Patent Document 1 does not disclose a specific analysis method and an abnormality detection method for time-series data such as acceleration data. The vibration waveform of the movable part is complex, and it is not always easy to determine an abnormality in the injection molding machine.

[0005] An object of the present disclosure is to provide a computer program, an abnormality detection method, an abnormality detection device, a molding machine system, and a learning model generation method that can convert time-series data of a physical quantity related to a movable part of an industrial machine into a time-series data image represented by an image and determine an abnormality in the industrial machine based on the time-series data image.

Means for Solving the Problems

[0006] The computer program relating to this disclosure is a computer program for causing a computer to perform a process to detect abnormalities in an industrial machine having a movable part, and involves acquiring time-series physical quantity data output from a sensor that detects physical quantities related to the motion of the movable part, converting the acquired time-series physical quantity data into a time-series data image represented as an image, inputting the converted time-series data image into a learning model that has learned the characteristics of the time-series data image relating to a normal movable part, calculating the feature quantities of the time-series data image, and causing the computer to perform a process to determine whether or not there is an abnormality in the industrial machine based on the calculated feature quantities.

[0007] The anomaly detection method relating to this disclosure is an anomaly detection method for detecting anomalies in an industrial machine having a movable part, which involves acquiring time-series physical quantity data output from a sensor that detects physical quantities related to the motion of the movable part, converting the acquired time-series physical quantity data into a time-series data image represented as an image, inputting the converted time-series data image into a learning model that has learned the characteristics of a time-series data image relating to a normal movable part, thereby calculating the feature quantities of the time-series data image, and determining whether or not there is an anomaly in the industrial machine based on the calculated feature quantities.

[0008] The abnormality detection device according to this disclosure comprises the abnormality detection device and a molding machine, wherein the abnormality detection device is configured to detect abnormalities in the molding machine.

[0009] An anomaly detection device according to this disclosure is an anomaly detection device for detecting anomalies in an industrial machine having a movable part, and comprises: a sensor for detecting physical quantities related to the motion of the movable part; an acquisition unit for acquiring time-series physical quantity data output from the sensor; a conversion unit for converting the time-series physical quantity data acquired by the acquisition unit into a time-series data image represented as an image; a calculation unit for calculating feature quantities of the time-series data image by inputting the converted time-series data image into a learning model that has learned the features of the time-series data image relating to a normal movable part; and a determination unit for determining whether or not there is an anomaly in the industrial machine based on the feature quantities calculated by the calculation unit.

[0010] The learning model generation method relating to this disclosure is a method for generating a learning model for detecting abnormalities in a twin-screw compounding extruder having a first screw and a second screw, and comprises a first image that plots the displacements in the first and second axial directions intersecting the rotational axis of the first screw of the twin-screw compounding extruder, based on time-series data showing the displacements in the third and fourth axial directions intersecting the rotational axis of the second screw of the twin-screw compounding extruder during normal operation, and The computer generates multiple time-series data images, each containing a second image that depicts the displacements in the third and fourth axis directions intersecting the rotational axis of the second screw. Based on the generated multiple time-series data images and a training dataset containing multiple reference images with arbitrary features, the computer executes a process to generate a learning model that outputs feature quantities corresponding to the normal and abnormal operation of the twin-screw compounding extruder when a time-series data image containing a first image and a second image depicting the displacements of the rotational axes of the first and second screws is input. [Effects of the Invention]

[0011] According to this disclosure, time-series data of physical quantities related to the moving parts of industrial machinery can be converted into a time-series data image, and abnormalities in the industrial machinery can be determined based on the said time-series data image. [Brief explanation of the drawing]

[0012] [Figure 1] This is a block diagram showing an example configuration of the molding machine system according to this embodiment. [Figure 2] This is a schematic diagram showing an example of the configuration of a twin-screw compounding extruder according to this embodiment. [Figure 3] This is a schematic diagram showing the mounting position of the displacement sensor. [Figure 4] This is a block diagram showing an example configuration of the anomaly detection device according to this embodiment. [Figure 5] This flowchart shows the anomaly detection processing procedure according to this embodiment. [Figure 6] It is a first time-series data image showing the displacement of the first screw. [Figure 7] It is a time-series data image showing the displacements of the first and second screws. [Figure 8] It is a conceptual diagram showing the feature amounts of the time-series data image and the anomaly detection results. [Figure 9] It is a conceptual diagram showing an overview of the method for generating a learning model. [Figure 10] It is a conceptual diagram showing the learning model in the learning phase. [Figure 11] It is a conceptual diagram showing the learning model in the test phase. [Figure 12] It is a scatter diagram showing the feature amounts of the time-series data image. [Figure 13] It is a chart showing the evaluation results of anomaly detection using the learning model.

Embodiments for Carrying Out the Invention

[0013] Specific examples of the computer program, anomaly detection method, anomaly detection device, molding machine system, and learning model generation method according to the embodiments of the present invention will be described below with reference to the drawings. Note that the present invention is not limited to these examples, and is defined by the claims, and is intended to include all modifications within the meaning and scope equivalent to the claims. Also, at least a part of the embodiments and modification examples described below may be arbitrarily combined.

[0014] FIG. 1 is a block diagram showing a configuration example of a molding machine system according to the present embodiment. The molding machine system includes a twin-screw kneading extruder 1, a displacement sensor 2, and an anomaly detection device 3.

[0015] <Twin-screw kneading extruder 1> Figure 2 is a schematic diagram showing a configuration example of the twin-screw kneading extruder 1 according to the present embodiment. The twin-screw kneading extruder 1 includes a cylinder 10 having a hopper 10a into which a resin raw material is charged, a first screw 11, and a second screw 12. The first and second screws 11, 12 are arranged substantially parallel to each other in a meshed state, and are rotatably inserted into the hole of the cylinder 10, and convey the resin raw material charged into the hopper 10a in the extrusion direction (right direction in FIG. 2), and melt and knead it. The first screw 11 is configured as a single screw by combining and integrating a plurality of types of screw pieces. For example, a forward flight piece having a flight screw shape for transporting the resin raw material in the forward direction, a reverse flight piece for transporting the resin raw material in the reverse direction, a kneading piece for kneading the resin raw material, etc. are arranged and combined in an order and position according to the characteristics of the resin raw material, whereby the first screw 11 is configured. The configuration of the second screw 12 is the same as that of the first screw 11.

[0016] Further, the twin-screw kneading extruder 1 includes a motor 13 that outputs a driving force for rotating the first and second screws 11, 12, and a speed reducer 14 that decelerates and transmits the driving force of the motor 13. The drive shafts 11a, 12a of the first and second screws 11, 12 are connected to the output shaft of the speed reducer 14. The first and second screws 11, 12 rotate by the driving force of the motor 13 decelerated and transmitted by the speed reducer 14.

[0017] <Displacement sensor 2> The displacement sensor 2 is a sensor that detects the displacement of the rotation center axes of the rotating first and second screws 11, 12. The displacement sensor 2 is preferably arranged at an appropriate position between the speed reducer 14 and the cylinder 10, specifically, in the vicinity of the drive shafts 11a, 12a of the first and second screws 11, 12. The first and second screws 11, 12 are an example of a movable part of an industrial machine. More specifically, the first and second screws 11, 12 are an example of a rotating shaft of a twin-screw kneading extruder 1 which is an example of an industrial machine.

[0018] Figure 3 is a schematic diagram showing the mounting position of the displacement sensor 2. The displacement sensor 2 comprises a first displacement sensor 21 that detects displacement in the X-axis direction (first direction) perpendicular to the rotational axis of the first screw 11, and a second displacement sensor 22 that detects displacement in the Y-direction (second axis direction) perpendicular to the rotational axis of the first screw 11 and the X-axis. The first displacement sensor 21 and the second displacement sensor 22 are arranged to face the drive shaft 11a of the first screw 11. The X-axis direction and the Y-axis direction are, for example, the horizontal and vertical directions perpendicular to the rotational axis of the first screw 11. Furthermore, the displacement sensor 2 includes a third displacement sensor 23 that detects displacement in the X-axis direction (third direction) perpendicular to the rotational axis of the second screw 12, and a fourth displacement sensor 24 that detects displacement in the Y-direction (fourth axis direction) perpendicular to the rotational axis of the second screw 12 and the X-axis. The third displacement sensor 23 and the fourth displacement sensor 24 are arranged to face the drive shaft 12a of the second screw 12. The X-axis direction and the Y-axis direction are, for example, the horizontal and vertical directions perpendicular to the rotational axis of the second screw 12. The first to fourth displacement sensors 21, 22, 23, and 24 constantly detect the displacement of the rotational axis of the rotating first and second screws 11 and 12, and output time-series displacement data (physical quantity data) showing the detected displacement to the anomaly detection device 3. Hereinafter, the first to fourth displacement sensors 21, 22, 23, and 24 will be collectively referred to as displacement sensor 2. The displacement data output by the first and second displacement sensors 21 and 22 corresponds to time-series data representing the first physical quantity related to this disclosure, and this displacement data is a two-dimensional physical quantity consisting of two numerical values. The displacement data output by the third and fourth displacement sensors 23 and 24 corresponds to time-series data representing the second physical quantity related to this disclosure, and this displacement data is a two-dimensional physical quantity consisting of two numerical values.

[0019] <Anomaly detection device 3> Figure 4 is a block diagram showing an example configuration of the anomaly detection device 3 according to this embodiment. The anomaly detection device 3 is a computer and comprises a processing unit 31, a storage unit 32, an input interface 33 (input I / F), and an output interface 34 (output I / F). The storage unit 32, the input interface 33, and the output interface 34 are connected to the processing unit 31.

[0020] The processing unit 31 includes arithmetic processing circuits such as a CPU (Central Processing Unit), multi-core CPU, GPU (Graphics Processing Unit), GPGPU (General-purpose computing on graphics processing units), TPU (Tensor Processing Unit), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), and NPU (Neural Processing Unit), as well as internal storage devices such as ROM (Read Only Memory) and RAM (Random Access Memory), and I / O terminals. The processing unit 31 functions as an anomaly detection device 3 according to this embodiment by executing a computer program P stored in the storage unit 32, which will be described later. Note that each functional part of the anomaly detection device 3 may be implemented in software, or some or all of them may be implemented in hardware.

[0021] The storage unit 32 is a hard disk, EEPROM (Electrically Erasable Programmable Memory). The memory is a non-volatile memory such as ROM or flash memory. The storage unit 32 stores a computer program P for causing a computer to implement the anomaly detection method according to this embodiment, and a learning model 35. The learning model 35 is a convolutional neural network model that, when a time-series data image (described later) representing time-series data obtained by detecting the displacement of the first and second screws 11 and 12 is input, extracts features from the time-series data image and outputs the extracted feature quantities. The learning model 35 is a model having a feature extraction layer of a CNN (Convolutional Neural Network), for example, a one-class classification model. Details of the method and configuration of the learning model 35 will be described later.

[0022] The computer program P and learning model 35 according to this embodiment may be recorded on a recording medium 4 in a manner that is computer-readable. The storage unit 32 stores the computer program P and learning model 35 read from the recording medium 4 by a reading device (not shown). The recording medium 4 is a semiconductor memory such as flash memory. The recording medium 4 may also be an optical disc such as a CD (Compact Disc)-ROM, DVD (Digital Versatile Disc)-ROM, or BD (Blu-ray® Disc). Furthermore, the recording medium 4 may be a magnetic disc such as a flexible disk or hard disk, or a magneto-optical disc. In addition, the computer program P and learning model 35 according to this embodiment may be downloaded from an external server (not shown) connected to a communication network (not shown) and stored in the storage unit 32.

[0023] The input interface 33 is connected to the first to fourth displacement sensors 21, 22, 23, and 24, and receives displacement data, which is time-series data output from the first to fourth displacement sensors 21, 22, 23, and 24.

[0024] A display unit 3a is connected to the output interface 34. The display unit 3a displays whether or not there is a malfunction in the twin-screw compounding extruder 1.

[0025] <Anomaly detection processing> Figure 5 is a flowchart showing the abnormality detection processing procedure according to this embodiment. The processing unit 31 acquires time-series displacement data output from the displacement sensor 2 (step S11). Preferably, the processing unit 31 acquires displacement data over the time it takes for the first and second screws 11 and 12 to rotate multiple times.

[0026] Next, the processing unit 31 converts the acquired displacement data into a first time-series data image (first image) and a second time-series data image (second image), which represent the displacement of the rotational center axes of the first and second screws 11 and 12, respectively (step S12).

[0027] Figure 6 is the first time-series data image showing the displacement of the first screw 11. The first time-series data image is a roughly square image, and the X and Y axes of this image correspond to the X-axis direction (first axis direction) and Y-axis direction (second axis direction) shown in Figure 4, respectively. The displacement amounts in the X-axis and Y-axis directions are plotted as the coordinate values ​​of the X and Y axes. The first time-series data image plots the displacement for three rotations.

[0028] The second time-series data image showing the displacement of the second screw 12 is also a roughly square image, similar to the first time-series data image. The X and Y axes of this image correspond to the X-axis direction (third axis direction) and the Y-axis direction (fourth axis direction), respectively, as in Figure 4, and the displacement amounts in the X-axis and Y-axis directions are plotted as the coordinate values ​​of the X and Y axes.

[0029] Next, the processing unit 31 combines the first time-series data image and the second time-series data image into a single time-series data image (step S13).

[0030] Figure 7 is a time-series data image showing the displacement of the first and second screws 11 and 12. The combined time-series data image is a roughly square image with twice the size (four times the area) of the first time-series data image. The time-series data image contains the information of the original images without changing the aspect ratio of the first and second time-series data images. Specifically, the time-series data image has the first and second time-series data images arranged side by side on the upper left and right sides of Figure 7, and a blank image placed at the bottom to make the time-series data image roughly square. The arrangement of the first and second time-series data images and the blank image is not particularly limited; as long as the first and second time-series data images are included as they are, the arrangement method is not particularly limited.

[0031] Next, the processing unit 31 inputs the time-series data image synthesized in step S13 into the learning model 35 to calculate the features of the time-series data image (step S14). The features are represented, for example, by multi-dimensional vector quantities, such as hundreds to thousands of numerical values.

[0032] Then, the processing unit 31 calculates the outlier score of the time-series data image features calculated in step S14 (step S15). The outlier score is a numerical representation of the degree of outlierity of the time-series data image features calculated in step S14 compared to the features of the time-series data image obtained during the normal operation of the twin-screw compounding extruder 1 (hereinafter referred to as sample features). The outlier score is, for example, LOF (Local Outlier Factor). LOF is the local density ld(P) of the feature targeted for anomaly detection and the average local density of the neighboring group of that feature (the k nearest sample features to that feature).<ld(Q)> Ratio<ld(Q)> This can be expressed as / ld(P). The greater the LOF (Low-Output Factor) value, the higher the degree of outlier status. Note that P represents the feature targeted for anomaly detection, and Q represents the sample feature mentioned above.

[0033] Then, the processing unit 31 determines whether the twin-screw compounding extruder 1 and the first and second screws 11 and 12 are functioning normally by determining whether the calculated outlier score is above a predetermined threshold (step S16). If the outlier score is the above LOF, the threshold is a value of 1 or greater. If the LOF is above the threshold, the processing unit 31 determines that there is an abnormality, and if the LOF is below the threshold, it determines that there is a normal condition.

[0034] The anomaly detection method using LOF described above is just one example; the twin-screw compounding extruder 1 may also be determined using the k-nearest neighbor method or SVM. Alternatively, the system may be configured to determine whether the twin-screw compounding extruder 1 is abnormal by using Hotelling's method to determine whether the features of the anomaly detection target are significantly different from the sample features under normal conditions.

[0035] Figure 8 is a conceptual diagram showing the features and anomaly detection results of a time-series data image. Figure 8 plots the high-dimensional features of a time-series data image on a two-dimensional plane after dimensionality reduction to two-dimensional features. The horizontal and vertical axes of Figure 8 represent the first and second features of the dimensionality-reduced time-series data image. Note that dimensionality reduction of features can be performed using dimensionality reduction algorithms such as t-SNE. As shown in Figure 8, the learning model 35 is trained to have a high local density of normal features output from the learning model 35. The dashed circles represent the threshold. If the statistical distance between the feature vectors of the time-series data image targeted for anomaly detection (e.g., star-shaped hexagon plot) and the feature vector group of the time-series data image under normal conditions is short (local density is relatively high) and the LOF value is small, the twin-screw compounding extruder 1 is presumed to be in a normal state. If the statistical distance between the feature vectors of the time-series data image targeted for anomaly detection (e.g., X-mark plot) and the feature vector group of the time-series data image under normal conditions is long (local density is relatively low) and the LOF value is large, the twin-screw compounding extruder 1 is presumed to be in an abnormal state.

[0036] If the twin-screw compounding extruder 1 is determined to be normal (step S16: YES), the processing unit 31 terminates processing. If the twin-screw compounding extruder 1 is determined to be abnormal (step S16: NO), the processing unit 31 outputs a warning that there is an abnormality in the twin-screw compounding extruder 1 (step S17) and terminates processing. For example, the processing unit 31 displays a warning screen on the display unit 3a to notify that there is an abnormality in the twin-screw compounding extruder 1. Alternatively, as shown in Figure 8, the processing unit 31 may be configured to display an image on the display unit 3a that represents the feature quantities of the current time-series data image being evaluated and the feature quantities of the time-series data when it is normal. Specifically, the processing unit 31 can reduce each high-dimensional feature quantity to two dimensions using an arbitrary dimensionality reduction algorithm such as t-SNE, and generate an image representing the feature quantities by plotting each two-dimensionally represented feature quantity on a two-dimensional graph. The processing unit 31 then outputs the image data to the display unit 3a, thereby displaying an image on the display unit 3a that shows the feature quantities of the current time-series data image being evaluated and the feature quantities of the time-series data under normal conditions.

[0037] <Method for generating a learning model> Figure 9 is a conceptual diagram showing an overview of the method for generating the learning model 35. Below, we will explain an example in which the processing unit 31 of the anomaly detection device 3 uses machine learning to train the learning model 35.

[0038] First, to generate a one-class classification model, we prepare a training dataset consisting of multiple time-series data images obtained when the twin-screw compounding extruder 1 is operating normally, and several arbitrary reference images R0, R1, and R2 that are unrelated to the time-series data images. The reference images R0, R1, and R2 are multi-class image data selected from a training dataset such as ImageNet. In the example shown in Figure 9, reference images R0, R1, and R2 labeled with labels 0, 1, and 2 are provided.

[0039] Then, the learning model 35, which has a feature extraction layer consisting of multiple convolutional and pooling layers that extract image features, is trained using time-series data images from normal operation and reference images R0, R1, and R2. Specifically, the processing unit 31 trains the learning model 35 to output feature quantities that can distinguish the features of each image, that is, feature quantities that have a high local density of feature quantities in the time-series data images, as shown in the right figure of Figure 9, and that also have high discriminability with the feature quantities of the reference images R0, R1, and R2. The following details how to generate a one-class classification model.

[0040] Figure 10 is a conceptual diagram showing the learning model 35 in the learning phase. First, a first neural network (reference network) 35a and a second neural network (secondary network) 35a are prepared.

[0041] The first neural network 35a is a CNN having an input layer, a feature extraction layer, and a classification layer. The feature extraction layer has a repeating structure of multiple convolutional layers and pooling layers. The classification layer has, for example, one or more fully connected layers. The first neural network 35a is pre-trained using a training dataset such as ImageNet. The second neural network 35a has the same network configuration as the first neural network 35a, and the various parameters (weight coefficients) that characterize the feature extraction layer and classification layer are also the same. Hereafter, in the learning phase of the first and second neural networks 35a, 35a, the various parameters of the feature extraction layer and classification layer share the same values.

[0042] The processing unit 31 then inputs a reference image R0 to the first neural network 35a and calculates a descriptive loss, which is a loss function. Meanwhile, the processing unit 31 inputs a normal time-series data image to the second neural network 35a and calculates a compactness loss, which is a loss function. The descriptive loss is a loss function commonly used in classifier training, such as the cross-entropy error. The compactness loss is expressed by the following equations (1), (2), and (3). The compactness loss is a value that corresponds to the variance of the output of the second neural network 35a within a batch, that is, in the training dataset.

[0043]

number

[0044] The processing unit 31 then calculates the descriptive loss and the total loss based on the compact loss. The total loss is expressed, for example, by the following equation (4).

[0045]

number

[0046] The processing unit 31 trains the first and second neural networks 35a, 35a by optimizing various parameters of the first and second neural networks 35a, 35a using methods such as backpropagation so that the total loss described in (4) above is reduced. When performing machine learning, it is preferable to fix the parameters of the preceding layers and adjust the parameters of the subsequent multiple layers.

[0047] As shown in Figure 9, the feature quantities output from the feature extraction layers of the first and second neural networks 35a, 35a, which have been trained in this manner, show a high local density of feature quantities in the time-series data images obtained when the twin-screw compounding extruder 1 is functioning normally, and a low local density with respect to feature quantities in other arbitrary images.

[0048] Figure 11 is a conceptual diagram showing the learning model 35 in the test phase. As shown in Figure 11, the learning model 35 according to this embodiment can be constructed using the input layer and feature extraction layer of the second neural network 35a trained as described above. Therefore, the classification layer shown in Figure 10 is not an essential component of the learning model 35 according to this embodiment. However, the second neural network 35a shown in Figure 10 can be used as the learning model 35 as is, and the features output from the feature extraction layer can be used. Note that Figure 11 shows two learning models 35, but this is a conceptual representation of a state where the time-series data image to be detected and the normal time-series data image (a sample) are input, and does not indicate that two models exist.

[0049] In the test phase, the processing unit 31 inputs time-series data images obtained under normal conditions as sample data into the learning model 35, thereby outputting the feature quantities of the time-series data images under normal conditions. On the other hand, it also inputs time-series data images targeted for anomaly detection into the learning model 35, thereby outputting the feature quantities of those time-series data images. The processing unit 31 then calculates an outlier score for the feature quantities of the target for anomaly detection relative to the feature quantities of the sample data, and by comparing the calculated outlier score with a threshold, it determines whether the time-series data image targeted for anomaly detection is abnormal, that is, whether the twin-screw compounding extruder 1 is abnormal.

[0050] In Figure 11, an example is shown in which sample data is input into the learning model 35 to calculate the features of the sample data. However, the features of multiple sample data sets may be calculated in advance and stored in the storage unit 32. In the anomaly detection process, the outlier score may be calculated using the features of the sample data stored in the storage unit 32.

[0051] Figure 12 is a scatter plot showing the features of time-series data images. Similar to Figure 8, Figure 12 is a scatter plot obtained by reducing the dimensionality of high-dimensional features of time-series data images to 2-dimensional features and plotting them on a 2-dimensional plane. The scatter plot shows feature A of the time-series data images obtained during the normal operation of the twin-screw compounding extruder 1, which is the training data; feature B of the normal time-series data images, which is the test data; and feature C of the abnormal time-series data images, which is the test data. As can be seen from Figure 12, feature B of the test data (normal) is in the vicinity of feature A of the training data, and feature C of the test data (abnormal) is separated from feature A of the training data in a discriminable manner.

[0052] Figure 13 is a chart showing the evaluation results of anomaly detection using the learning model 35. In the table in Figure 13, "Simulated Anomaly" indicates the type of anomaly. Wear indicates an abnormal state in which the first and second screws 11 and 12 are worn. In this experiment, the wear state was simulated by using an FF (forward lead full flight) piece before the kneading section and the first FK (forward shift kneading disc) in the kneading section with reduced outer diameters, and the displacement of the first and second screws 11 and 12 was detected to create a time-series data image. Specifically, when the screw outer diameter is 68.0 mm, the wear state was simulated using the above pieces with an outer diameter of 63.0 mm. Screw mismatch indicates an abnormal state in which the twin-screw kneading extruder 1 is operated using a different type of resin raw material than the assumed resin raw material. In this experiment, it shows an abnormal state in which nylon (PA) was used as the resin raw material with the first and second screws 11 and 12 for polypropylene (PP). The "normal data" shows time-series data images obtained when the first and second screws 11 and 12 were not worn, and when the first and second screws 11 and 12 were rotated at rotational speeds of 100 rpm, 200 rpm, and 400 rpm using appropriate resin material (PP). The "abnormal data" section shows time-series data images obtained when the first and second screws 11 and 12 were rotated at rotational speeds of 100 rpm, 200 rpm, and 400 rpm using the appropriate resin material (PP) in the above-mentioned abnormal wear condition for "wear". The "screw non-conformity" section shows time-series data images obtained when the first and second screws 11 and 12 were not worn and were rotated at rotational speeds of 100 rpm, 200 rpm, and 400 rpm using the above-mentioned unsuitable resin material (PA).

[0053] Then, a portion of the normal data shown in the table in Figure 13 was used as the training dataset to train the learning model 35, and the ability to distinguish between the remaining normal data and the abnormal data was evaluated using AUC (Area Under the Curve). The AUCs for each of the normal and abnormal data were 0.96, 1.00, 0.99, 1.00, 1.00, and 0.94, confirming that the model possessed high discriminative ability.

[0054] As described above, according to the abnormality detection device 3, etc. of this embodiment, time-series data of the displacement of the first and second screws 11 and 12 of the twin-screw compounding extruder 1 can be converted into a time-series data image, and an abnormality of the twin-screw compounding extruder 1 can be determined based on the time-series data image.

[0055] Furthermore, by configuring the learning model 35 as a single-class classification model, it is possible to train the learning model 35 using time-series data images obtained during normal operation of the twin-screw compounding extruder 1, without using time-series data images of the twin-screw compounding extruder 1 during abnormal conditions.

[0056] Furthermore, by representing the time-series changes in the X-axis and Y-axis displacements of the first and second screws 11 and 12 in a single time-series data image, and calculating the feature quantities of this time-series data image using the learning model 35, it is possible to easily obtain feature quantities that accurately represent the state of the first and second screws 11 and 12 of the twin-screw compounding extruder 1. By using the feature quantities obtained in this way, abnormalities in the twin-screw compounding extruder 1 can be detected with high accuracy.

[0057] Furthermore, by using a time-series data image that encompasses the first and second time-series data images representing the displacements of the first and second screw shafts 11 and 12 respectively without deformation, abnormalities in the twin-screw compounding extruder 1 can be detected with high accuracy. It has been confirmed that using images obtained by stretching the first and second time-series data images vertically or horizontally affects the characteristics of the first and second screws 11 and 12, resulting in a decrease in anomaly detection accuracy.

[0058] In this embodiment, an example was described in which displacement sensors 2 are provided on the drive shafts 11a and 12a of the first and second screws 11 and 12. However, displacement sensors 2 may be provided on other parts as long as it is possible to detect abnormalities in the first and second screws 11 and 12. For example, displacement sensors 2 may be provided on the output shaft of the motor 13. The output shaft is an example of a rotating shaft of a molding machine.

[0059] Furthermore, although this embodiment describes an example in which displacement sensors 2 are provided on each of the first and second screws 11 and 12, the system may also be configured to detect abnormalities by providing displacement sensors 2 on only one of the first and second screws 11 and 12.

[0060] Furthermore, although this embodiment describes an example of detecting the displacement of the rotational axis of the first and second screws 11 and 12, the twin-screw compounding extruder 1 may be configured to detect physical quantities such as torque, rotational speed, or rotational acceleration applied to the first and second screws 11 and 12, and to use the time-series data of said physical quantities. Alternatively, the twin-screw compounding extruder 1 may be configured to detect abnormalities using a time-series data image that includes multiple images depicting different types of physical quantities. For example, the twin-screw compounding extruder 1 may be configured to detect abnormalities using a time-series data image that includes a first time-series data image depicting the displacement of the rotational axis of the first and second screws 11 and 12, and a time-series data image depicting the torque or rotational acceleration applied to the first and second screws 11 and 12. Similarly, the system may be configured to detect abnormalities in industrial machinery by detecting any physical quantity in the movable parts.

[0061] Furthermore, in this embodiment, an example was described in which a time-series data image is used that includes a two-dimensional first time-series data image (first image) showing the displacement of the first screw 11 with two numerical values, and a two-dimensional second time-series data image (second image) showing the displacement of the second screw 12 with two numerical values. However, the system may also be configured to detect abnormalities in the twin-screw compounding extruder 1 using a time-series data image that is a composite of four images depicting four types of two-dimensional physical quantity data. Alternatively, the system may be configured to detect abnormalities in the twin-screw compounding extruder 1 using a time-series data image that includes five or more images depicting time-series data.

[0062] Furthermore, although this embodiment describes an example in which a time-series data image is used, which is obtained by directly plotting a physical quantity that changes over time, it is also possible to configure the system to detect abnormalities in the twin-screw compounding extruder 1 or industrial machinery by frequency-converting the detected time-series physical quantity and plotting the frequency spectrum of the physical quantity in a time-series data image.

[0063] Furthermore, although this embodiment describes an example of detecting abnormalities in the first and second screws 11 and 12 of the twin-screw compounding extruder 1, the abnormality detection device 3 may also be configured to detect abnormalities in other movable parts, such as single-screw extruders, injection molding machines, and other industrial machinery. In this case, the learning model 35 learns time-series data images of the movable part when it is functioning normally. In the abnormality detection device 3 for detecting abnormalities in the moving parts of industrial machinery, the processing unit 31 acquires time-series data representing a two-dimensional first physical quantity consisting of two numerical values ​​and time-series data representing a two-dimensional second physical quantity consisting of two numerical values, similar to the embodiment. The acquired first time-series data is converted into a first image depicting the first physical quantity, and the acquired second time-series data is converted into a second image depicting the second physical quantity. The time-series data image is then synthesized, containing the converted first and second images as a single image. The processing unit 31 then inputs the synthesized time-series data image into a learning model to calculate the feature quantities of the time-series data image and calculates an outlier score for the feature quantities of the normal time-series data image. The processing unit 31 can detect abnormalities in industrial machinery by comparing the outlier score with a predetermined threshold. Furthermore, the system may be configured to detect abnormalities in industrial machinery using a time-series data image that includes three or more images depicting three or more different physical quantities. The method for generating the learning model 35 for detecting abnormalities in the moving parts of industrial machinery is the same as in the embodiment. It is preferable to train a one-class classification model using the method shown in Figure 10, with multiple time-series data images generated during the normal operation of the industrial machinery and multiple reference images as the training dataset.

[0064] Furthermore, although this embodiment describes an example in which the processing unit 31 of the anomaly detection device 3 performs machine learning on the learning model 35, the learning model 35 may also be performed using another external computer or server.

[0065] Furthermore, while a one-class classification model was described as an example of learning model 35, other neural networks such as general CNNs, U-Nets, and RNNs (Recurrent Neural Networks), as well as other SVMs (Support Vector Machines), Bayesian networks, or regression trees may also be used.

[0066] Furthermore, the system may be configured to distinguish typical anomaly time-series image data from other anomaly time-series image data using multiple single-class classification models, namely the learning models 35. For example, the anomaly detection device 3 includes a first learning model trained on normal time-series image data, similar to the embodiment described above. The anomaly detection device 3 also includes a second learning model 35 trained on normal time-series image data and a first anomaly time-series image. By using the first learning model 35 and the second learning model 35, the anomaly detection device 3 can distinguish between normal time-series image data, a first anomaly time-series image data, and other anomaly time-series image data. Similarly, by providing the anomaly detection device 3 with three or more learning models 35, it can be configured to distinguish between two or more anomaly time-series image data. Alternatively, the system may be configured to classify normal time-series image data from anomaly time-series image data using a machine learning-trained multi-class classification model.

[0067] Furthermore, while we have primarily described examples of machine learning using neural networks such as CNNs, the structure of the trained model is not limited to CNNs, RNNs, etc. It may also be constructed using other neural networks, SVMs (Support Vector Machines), Bayesian networks, or regression trees. The means of this disclosure are noted below. (Note 1) A computer program for causing a computer to perform a process to detect abnormalities in industrial machinery having moving parts, Time-series physical quantity data is obtained from a sensor that detects the physical quantity related to the movement of the movable part. The acquired time-series physical quantity data is converted into a time-series data image, which represents the data as an image. By inputting the converted time-series data image into a learning model that has learned the features of the time-series data image relating to the normal movable part, the feature quantities of the time-series data image are calculated. Based on the calculated feature quantities, determine whether or not there is a malfunction in the industrial machine. A computer program that causes the computer to perform a process. (Note 2) The aforementioned learning model is a model obtained through machine learning. The computer program described in Appendix 1. (Note 3) The aforementioned learning model is a one-class classification model obtained through machine learning. The computer program described in Appendix 1 or Appendix 2. (Note 4) The aforementioned physical quantity data includes time series data representing a first physical quantity and time series data representing a second physical quantity. The aforementioned time-series data image is The first image, which depicts the first physical quantity, and the second image, which depicts the second physical quantity, are included in a single image. A computer program described in any one of the appendices 1 through 3. (Note 5) The first and second physical quantities are each two-dimensional physical quantities consisting of two numerical values. The computer program described in Appendix 4. (Note 6) The movable part is the rotating shaft of the molding machine, and the physical quantity data includes time-series data showing the displacement, torque, rotational speed, or rotational acceleration of the rotating shaft of the molding machine. A computer program described in any one of the appendices 1 through 5. (Note 7) The movable part is the screw of the molding machine, and the physical quantity data includes time-series data showing the displacement, torque, rotational speed, or rotational acceleration of the screw of the molding machine. A computer program described in any one of the appendices 1 through 6. (Note 8) The movable parts are the first and second screws of a twin-screw compounding extruder, and the physical quantity data includes time-series data showing the displacement in the first and second axial directions intersecting the rotational axis of the first screw, and time-series data showing the displacement in the third and fourth axial directions intersecting the rotational axis of the second screw. The aforementioned time-series data image is The image includes a first image that plots the displacement of the first screw using the intersecting first and second axis directions as coordinate axes, and a second image that plots the displacement of the second screw using the intersecting third and fourth axis directions as coordinate axes. A computer program described in any one of the appendices 1 through 3. (Note 9) The aforementioned time-series data image is a roughly square image. The first image and the second image, and a blank image that fills in the parts other than the first image and the second image. The computer program described in Appendix 8. (Note 10) An anomaly detection method for detecting abnormalities in industrial machinery having movable parts, Time-series physical quantity data is obtained from a sensor that detects the physical quantity related to the movement of the movable part. The acquired time-series physical quantity data is converted into a time-series data image, which represents the data as an image. By inputting the converted time-series data image into a learning model that has learned the features of the time-series data image relating to the normal movable part, the feature quantities of the time-series data image are calculated. Based on the calculated feature quantities, determine whether or not there is a malfunction in the industrial machine. Anomaly detection method. (Note 11) An anomaly detection device for detecting abnormalities in industrial machinery having movable parts, A sensor that detects a physical quantity related to the movement of the movable part, An acquisition unit that acquires time-series physical quantity data output from the sensor, A conversion unit converts the time-series physical quantity data acquired by the acquisition unit into a time-series data image represented as an image, A calculation unit calculates the feature quantities of the time-series data image by inputting the converted time-series data image into a learning model that has learned the features of the time-series data image relating to the normal movable part, A determination unit determines whether or not there is an abnormality in the industrial machine based on the feature quantities calculated by the calculation unit. An anomaly detection device equipped with the following features. (Note 12) An anomaly detection device as described in Appendix 11, Molding machine and Equipped with, The abnormality detection device is configured to detect abnormalities in the molding machine. Molding machine system. (Note 13) A method for generating a learning model for detecting abnormalities in a twin-screw compounding extruder having a first screw and a second screw, Based on time-series data showing the displacements in the first and second axial directions intersecting the rotational axis of the first screw of the twin-screw compounding extruder during normal operation, and time-series data showing the displacements in the third and fourth axial directions intersecting the rotational axis of the second screw of the twin-screw compounding extruder during normal operation, multiple time-series data images are generated, each containing a first image depicting the displacements in the first and second axial directions intersecting the rotational axis of the first screw, and a second image depicting the displacements in the third and fourth axial directions intersecting the rotational axis of the second screw. Based on a training dataset that includes the generated multiple time-series data images and multiple reference images having arbitrary features, a learning model is generated that outputs feature quantities corresponding to the normal and abnormal operation of the twin-screw compounding extruder when time-series data images including the first and second images depicting the displacement of the rotational center axes of the first and second screws are input. A method for generating learning models in which a computer performs the processing. [Explanation of symbols]

[0068] 1. Twin-screw compounding extruder 2. Displacement Sensor 3. Anomaly detection device 3a Display section 4. Recording media 10 cylinders 10a hopper 11 First screw 11a Drive shaft 12. Second screw 12a Drive shaft 13 Motors 14 Reducer 21 First displacement sensor 22 Second displacement sensor 23 Third displacement sensor 24. Fourth displacement sensor 31 Processing Unit 32 Storage section 33 Input Interfaces 34 Output Interfaces 35 Learning Models 35a Neural Network P Computer Program

Claims

1. A computer program for causing a computer to perform a process to detect an abnormality in a twin-screw compounding extruder having a first screw and a second screw, The system acquires time-series physical quantity data output from sensors that detect physical quantities related to the motion of the first and second screws, which includes time-series data showing the time change in displacement in the first axial direction and the second axial direction intersecting the rotational center axis of the first screw, and time-series data showing the time change in torque of the first and second screws. Based on the acquired time-series physical quantity data, a time-series data image is created that includes, in a single image, a first image plotting the displacement of the first screw in the first and second directions at each time point, using intersecting X and Y directions corresponding to the first and second directions as coordinate axes, and a second image plotting the torque of the first and second screws at each time point, using intersecting second X and Y directions as coordinate axes. By inputting the converted time-series data image into a learning model that has learned the features of the time-series data image relating to the normal first and second screws, the feature quantities of the time-series data image are calculated. Based on the calculated feature quantities, determine whether or not there is an abnormality in the twin-screw extruder. A computer program that causes the computer to perform a process.

2. The aforementioned learning model is a model obtained through machine learning. The computer program according to claim 1.

3. The aforementioned learning model is a one-class classification model obtained through machine learning. The computer program according to claim 1 or claim 2.

4. The aforementioned physical quantity data further includes time-series data indicating the rotational speed or rotational acceleration of the first and second screws. A computer program according to any one of claims 1 to 3.

5. An abnormality detection method for detecting an abnormality in a twin-screw kneading extruder having a first screw and a second screw, The system acquires time-series physical quantity data output from sensors that detect physical quantities related to the motion of the first and second screws, which includes time-series data showing the time change in displacement in the first axial direction and the second axial direction intersecting the rotational center axis of the first screw, and time-series data showing the time change in torque of the first and second screws. Based on the acquired time-series physical quantity data, a time-series data image is created that includes, in a single image, a first image plotting the displacement of the first screw in the first and second directions at each time point, using intersecting X and Y directions corresponding to the first and second directions as coordinate axes, and a second image plotting the torque of the first and second screws at each time point, using intersecting second X and Y directions as coordinate axes. By inputting the converted time-series data image into a learning model that has learned the features of the time-series data image relating to the normal first and second screws, the feature quantities of the time-series data image are calculated. Based on the calculated feature quantities, determine whether or not there is an abnormality in the twin-screw extruder. Anomaly detection method.

6. An abnormality detection device for detecting an abnormality in a twin-screw kneading extruder having a first screw and a second screw, A sensor for detecting physical quantities related to the motion of the first screw and the second screw, An acquisition unit that acquires physical quantity data including time-series data of physical quantities output from the sensor, which includes time-series data showing the time change in displacement in the first axial direction and the second axial direction intersecting the rotational center axis of the first screw, and time-series data showing the time change in torque of the first screw and the second screw. Based on the time-series physical quantity data acquired by the acquisition unit, a conversion unit converts the data into a time-series data image that includes, in a single image, a first image which plots the displacement of the first screw in the first and second directions at each time point, using intersecting X and Y directions corresponding to the first and second directions as coordinate axes, and a second image which plots the torque of the first and second screws at each time point, using intersecting second X and Y directions as coordinate axes. A calculation unit calculates the feature quantities of the time-series data image by inputting the converted time-series data image into a learning model that has learned the features of the time-series data image relating to the normal first screw and second screw, A determination unit determines whether or not there is an abnormality in the twin-screw compounding extruder based on the characteristic quantities calculated by the calculation unit. An anomaly detection device equipped with the following features.

7. An anomaly detection device according to claim 6, Twin-screw compounding extruder and Equipped with, The abnormality detection device is configured to detect abnormalities in the twin-screw compounding extruder. Molding machine system.

Images