State estimation method, state estimation device, program product, and machine learning method

Abstract

The invention relates to a state estimation method, a state estimation device, a program product, and a machine learning method. The state estimation method determines the state of an extruder by acquiring physical quantity data associated with the state of the extruder and inputting the acquired physical quantity data or a feature quantity based on the physical quantity data to a machine learning model, the machine learning model outputs information indicating the state of the extruder when the physical quantity data or a feature quantity based on the physical quantity data is input.

Description

Technical Field

[0001] The present invention relates to a state estimation method, a state estimation device, a program product, and a machine learning method. Background Art

[0002] The following technique is disclosed in Japanese Unexamined Patent Application Publication No. 2009-131965: A load detection device provided at an input shaft portion of a screw shaft is used to detect a load (torque) on the input shaft portion. When the detected average load value and load amplitude value are not within a predetermined normal range, it is determined that an overload state exists. When the overload duration for maintaining the overload state exceeds a set time, an abnormal alarm is reported and / or the rotation of the screw shaft is stopped. Summary of the Invention

[0003] However, as a problem peculiar to an extruder, there is a problem that the range (scale) of detection values varies depending on the model. In addition, even for the same model, there is a problem that the range of detection values varies depending on the operating method and the difference in raw materials. If the raw materials and operating methods are different, it is difficult to perform abnormal detection with a simple threshold value. Usually, there is no clear criterion for the threshold value used for abnormal detection, and it is determined by a skilled person with driving experience, for example.

[0004] An object of the present disclosure is to provide a state estimation method, a state estimation device, a program product, and a machine learning method that can determine the presence or absence or degree of abnormality of an extruder regardless of the model, operating method, and difference in raw materials of the extruder.

[0005] Means for Solving the Problem

[0006] A state estimation method according to an aspect of the present disclosure determines the state of an extruder by acquiring physical quantity data associated with the state of the extruder and inputting the acquired physical quantity data or a feature quantity based on the physical quantity data to a machine learning model, and the machine learning model outputs information indicating the state of the extruder when the physical quantity data or the feature quantity based on the physical quantity data is input.

[0007] A state estimation device according to an aspect of the present disclosure includes a processing unit that executes the following processing: determining the state of the extruder by acquiring physical quantity data associated with the state of the extruder and inputting the acquired physical quantity data or a feature quantity based on the physical quantity data to a machine learning model, and the machine learning model outputs information indicating the state of the extruder when the physical quantity data or the feature quantity based on the physical quantity data is input.

[0008] A computer program included in a program product according to an aspect of the present disclosure causes a computer to perform the following process: determining a state of an extruder by acquiring physical quantity data associated with the state of the extruder and inputting the acquired physical quantity data or a feature quantity based on the physical quantity data into a machine learning model, the machine learning model outputting information indicating the state of the extruder when the physical quantity data or the feature quantity based on the physical quantity data is input.

[0009] A machine learning method according to an aspect of the present disclosure acquires physical quantity data associated with the state of an extruder and generates a machine learning model based on the acquired physical quantity data, the machine learning model outputting information indicating the state of the extruder when the physical quantity data or a feature quantity based on the physical quantity data is input.

[0010] Advantageous Effects of the Invention

[0011] According to the present disclosure, it is possible to determine the presence or absence or degree of abnormality of an extruder regardless of differences in the type, operation method, and raw material of the extruder. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a block diagram showing a structural example of an extruder system according to Embodiment 1.

[0013] Figure 2 is a schematic diagram showing a structural example of an extruder according to Embodiment 1.

[0014] Figure 3 is a block diagram showing a structural example of a data collection device according to Embodiment 1.

[0015] Figure 4 is a block diagram showing a structural example of a state estimation device according to Embodiment 1.

[0016] Figure 5 is a conceptual diagram showing a first autoencoder in a learning stage.

[0017] Figure 6 is a conceptual diagram showing a first autoencoder in a detection stage.

[0018] Figure 7 is a conceptual diagram showing a second autoencoder in a detection stage.

[0019] Figure 8 is a conceptual diagram showing a third autoencoder in a detection stage.

[0020] Figure 9 is a conceptual diagram showing an example of a record layout of a collection data DB.

[0021] Figure 10It is a flowchart showing the learning process sequence of the autoencoder.

[0022] Figure 11 It is a flowchart showing the processing sequence of the state estimation of the extruder in Embodiment 1.

[0023] Figure 12 It is a conceptual diagram showing the feature quantity calculation model and the second autoencoder in Embodiment 2.

[0024] Figure 13 It is a conceptual diagram showing the state estimation method in Embodiment 3.

[0025] Figure 14 It is a block diagram showing a structural example of the state estimation device in Embodiment 4.

[0026] Figure 15 It is a flowchart showing the processing sequence of the state estimation of the extruder in Embodiment 4.

[0027] Explanation of reference numerals

[0028] 1: Extruder

[0029] 2: Sensor

[0030] 3: Data collection device

[0031] 4: Router

[0032] 5: State estimation device

[0033] 6: Terminal device

[0034] 10: Cylinder

[0035] 11: Screw

[0036] 12: Die

[0037] 13: Motor

[0038] 14: Reducer

[0039] 15: Control device

[0040] 50: Recording medium

[0041] 51: Processing unit

[0042] 52: Storage unit

[0043] 53: Communication unit

[0044] 54: First autoencoder

[0045] 55: Second autoencoder

[0046] 56: Third autoencoder

[0047] 57: Collect data DB

[0048] P: Computer program Detailed implementation manner

[0049] The following describes the state estimation method, state estimation device, program product, and machine learning method of the embodiments of the present disclosure with reference to the accompanying drawings. It should be noted that the present disclosure is not limited to these examples, as shown in the claims, and is intended to include all changes within the meaning and scope equivalent to the claims. In addition, at least a part of the embodiments described below can be arbitrarily combined.

[0050] Figure 1 It is a block diagram showing a structural example of the extruder system of Embodiment 1. The extruder system includes an extruder 1, a plurality of sensors 2, a data collection device 3, a router 4, a state estimation device 5, and a terminal device 6.

[0051] In Figure 1 One extruder 1 and a data collection device 3 are illustrated, but it can also be configured such that a plurality of data collection devices 3 are connected to the state estimation device 5 via a network. It can also be configured such that one or more extruders 1 are connected to the data collection device 3. The state estimation device 5 can collect information on each of one or more extruders 1 and estimate the state of each extruder 1.

[0052] The terminal device 6 is a communication terminal having a display unit, such as a computer, a tablet terminal, or a smartphone.

[0053] <Extruder 1>

[0054] Figure 2 It is a schematic diagram showing a structural example of the extruder 1 of Embodiment 1. The extruder 1 includes a cylinder 10, two screws 11, and a die 12 provided at the outlet portion of the cylinder 10 (refer to Figure 1 ). The cylinder 10 has an inlet 10a for injecting a resin raw material. The resin raw material is supplied from the inlet 10a to the cylinder 10 through a feeder 10b. The feeder 10b is a device that supplies the raw material to the cylinder 10 of the extruder while performing weight control. The two screws 11 are arranged substantially parallel to each other in a meshing state and are rotatably inserted into the hole of the cylinder 10, and convey the resin raw material injected into the inlet 10a in the extrusion direction ( Figure 1 and Figure 2 the right direction in ), and perform melting and kneading. The molten resin raw material is discharged from the die 12 having a through hole.

[0055] The screw 11 is formed as a single screw 11 by combining and integrating a plurality of screw members. For example, a forward thread member having a thread screw shape for conveying the resin raw material in the forward direction, a reverse thread member for conveying the resin raw material in the reverse direction, a kneading member for kneading the resin raw material, etc. are arranged and combined in an order and position corresponding to the characteristics of the resin raw material, thereby forming the screw 11.

[0056] In addition, the extruder 1 includes a motor 13 that outputs a driving force for rotating the screw 11, a speed reducer 14 that reduces and transmits the driving force of the motor 13, and a control device 15. The screw 11 is connected to the output shaft of the speed reducer 14. The screw 11 rotates by the driving force of the motor 13 that is reduced and transmitted by the speed reducer 14.

[0057] <Sensor 2>

[0058] The sensor 2 detects a physical quantity related to the state of the components constituting the extruder 1, and directly or indirectly outputs the detected physical quantity data to the data collection device 3. The physical quantity data is data representing the sensor values of the detected physical quantity in a time series. The sensor 2 includes sensors provided in the extruder 1 as sensors required for the operation control of the extruder 1, and sensors separately provided for estimating the state of the extruder 1. A part of the plurality of sensors 2 is connected to the data collection device 3, and the data collection device 3 obtains physical quantity data from the sensor 2. A part of the plurality of sensors 2 is connected to the control device 15, and the data collection device 3 obtains physical quantity data from the sensor 2 via the control device 15.

[0059] Examples of the physical quantity include temperature, position, speed, acceleration, current, voltage, pressure, time, image data, torque, force, strain, power consumption, weight, etc. These physical quantities can be measured using a thermometer, a position sensor, a speed sensor, an acceleration sensor, an ammeter, a voltmeter, a pressure gauge, a timer, a camera, a torque sensor, a power meter, a weighing scale, etc.

[0060] The plurality of sensors 2, for example, include a raw material supply amount detector, a screw speed detector, an extruder power detector, an extruder torque detector, an extruder current detector, an extruder load detector, a resin pressure detector, a resin temperature detector, a temperature detector for a cylinder or a barrel, an output detector for a cylinder or a barrel, an imaging unit that captures a static image or a dynamic image of an opening of the extruder 1, and a pressure sensor that detects the pressure at the discharge port of the extruder 1. The raw material supply amount detector is a sensor that detects the supply amount of the resin raw material input by the feeder. The extruder power detector is a detection sensor that detects the power of the motor 13. The extruder torque detector is a sensor that detects the torque acting on the motor 13 or the screw 11. In addition, the sensor 2 further includes any detector that can detect the throughput of the extruder 1 as an extruder and physical quantities that contribute to the estimation of the screw speed and load.

[0061] In addition, the physical quantity data output from the control device 15 to the data collection device 3 may also include the set value of the screw speed, the set value of the torque of the extruder 1, the set value of the current or load, the set value of the raw material supply amount, the set value of the temperature or output of the cylinder or barrel. In addition, the physical quantity data may also include static information (data representing static characteristics), such as the diameter of the extruder 1, the screw length, the load limit, the maximum speed, information representing the structure of the screw, and index information related to the screw structure.

[0062] <Control device 15>

[0063] The control device 15 is a computer that controls the operation of the extruder 1, and includes a transceiver unit (not shown) and a display unit that communicate information with the data collection device 3.

[0064] Specifically, the control device 15 sends operation data indicating the operation state of the extruder 1 to the data collection device 3. The operation data, for example, includes the raw material supply amount, the screw speed, the extruder power, the extruder torque, the extruder current, the resin pressure, the resin temperature, etc.

[0065] The control device 15 receives the state estimation result of the extruder 1 sent from the data collection device 3 and displays the received state estimation result. The state estimation result is information indicating the presence or absence and degree of abnormality of the extruder 1 and its components.

[0066] <Data collection device 3>

[0067] Figure 3FIG. 0 is a block diagram showing a structural example of the data collection device 3 according to Embodiment 1. The data collection device 3 is a computer, and includes a control unit 31, a storage unit 32, a communication unit 33, and a data input unit 34. The storage unit 32, the communication unit 33, and the data input unit 34 are connected to the control unit 31. The data collection device 3 is, for example, a PLC (Programmable Logic Controller).

[0068] The control unit 31 includes an arithmetic processing circuit such as a CPU (Central Processing Unit), a multi-core CPU, an ASIC (Application Specific Integrated Circuit), or an FPGA (Field-Programmable Gate Array), an internal storage device such as a ROM (Read Only Memory) or a RAM (Random Access Memory), and I / O terminals. The control unit 31 executes a process of collecting physical quantity data and transmitting it to the state estimation device 5 by executing a control program stored in the storage unit 32 described later. In addition, each functional unit of the data collection device 3 can be implemented by software or a part or all of it can be implemented by hardware.

[0069] The storage unit 32 is a non-volatile memory such as a hard disk, an EEPROM (Electrically Erasable Programmable ROM), or a flash memory. The storage unit 32 stores a control program for causing a computer to perform a process of collecting physical quantity data.

[0070] The communication unit 33 is a communication circuit that transmits and receives information according to a prescribed communication protocol such as Ethernet (registered trademark). The communication unit 33 is connected to the control device 15 via a first communication network such as a LAN. The control unit 31 can transmit and receive various information to and from the control device 15 via the communication unit 33. The control unit 31 acquires physical quantity data via the communication unit 33.

[0071] A router 4 is connected to the first network. The communication unit 33 is connected to the state estimation device 5 on the cloud as a second communication network via the router 4. The control unit 31 can transmit and receive various information to and from the state estimation device 5 via the communication unit 33 and the router 4.

[0072] The data input unit 34 is an input interface that inputs a signal output from the sensor 2. The sensor 2 is connected to the data input unit 34. The control unit 31 acquires physical quantity data via the data input unit 34.

[0073] <State estimation device 5>

[0074] Figure 4 This is a block diagram showing a structural example of the state estimation device 5 according to the first embodiment. The state estimation device 5 is a computer, and includes a processing unit 51, a storage unit 52, and a communication unit 53. The storage unit 52 and the communication unit 53 are connected to the processing unit 51.

[0075] The processing unit 51 is a processor, and has arithmetic processing circuits such as a CPU, a multi-core CPU, a GPU (Graphics Processing Unit), a GPGPU (General-purpose computing on graphics processing units), a TPU (Tensor Processing Unit), an ASIC, an FPGA, an NPU (Neural Processing Unit), internal storage devices such as a ROM and a RAM, I / O terminals, and the like. The processing unit 51 functions as the state estimation device 5 of the present embodiment by executing the computer program P (program product) stored in the storage unit 52 described later. In addition, each functional unit of the state estimation device 5 can be implemented by software, or a part or all of them can be implemented by hardware.

[0076] The communication unit 53 is a communication circuit that transmits and receives information according to a specified communication protocol such as Ethernet (registered trademark). The communication unit 53 is connected to the data collection device 3 and the terminal device 6 via a second communication network, and the processing unit 51 can transmit and receive various information to and from the data collection device 3 and the terminal device 6 via the communication unit 53.

[0077] The storage unit 52 is a non-volatile memory such as a hard disk, an EEPROM, or a flash memory. The storage unit 52 stores a computer program P for causing a computer to execute processing for estimating the presence or absence and the degree of abnormality of the extruder 1 and its constituent components, a first autoencoder 54, a second autoencoder 55, a third autoencoder 56, and a collected data DB57.

[0078] The computer program P etc. can also be in a form that can be read by a computer and recorded on the recording medium 50. The storage unit 52 stores the computer program P etc. read from the recording medium 50 by a reading device (not shown). The recording medium 50 is a semiconductor memory such as a flash memory. In addition, the recording medium 50 can also be an optical disc such as a CD (Compact Disc) - ROM, a DVD (Digital Versatile Disc) - ROM, or a BD (Blu - ray (registered trademark) Disc). Moreover, the recording medium 50 can also be a magnetic disc such as a floppy disk or a hard disk, or a magneto - optical disc. In addition, the computer program P etc. can be downloaded from an external server (not shown) connected to a communication network (not shown) and stored in the storage unit 52.

[0079] The first auto - encoder 54 is an auto - encoder that, when the physical quantity data detected by the sensor 2 is input, reproduces and outputs the physical quantity data that should be obtained from the normal extruder 1 regardless of the state at the time of detection of the physical quantity data.

[0080] Figure 5 It is a conceptual diagram showing the first auto - encoder 54 in the learning stage. Figure 6 It is a conceptual diagram showing the first auto - encoder 54 in the detection stage. When physical quantity data is input to the first auto - encoder 54, the first auto - encoder 54 extracts the features contained in the physical quantity data and outputs physical quantity data that reproduces the features of the physical quantity that can be obtained in the normal state. As Figure 6 shown, when the physical quantity data detected when the extruder 1 is abnormal is input to the first auto - encoder 54, the first auto - encoder 54 outputs physical quantity data that reproduces the features of the physical quantity that should be obtained when the extruder 1 is normal.

[0081] The first auto - encoder 54 has an input layer 54a, an intermediate layer 54b, and an output layer 54c. The intermediate layer 54b is a network structure that has an encoding layer and a decoding layer and is symmetric on the input side and the output side. The input layer 54a is the layer for inputting physical quantity data. In addition, as the physical quantity data, the detected value detected at the current moment and the detected values at multiple moments earlier than the current moment can be input to the input layer 54a. The encoding layer of the intermediate layer 54b is the layer for compressing the dimension of the physical quantity data. Through the dimension compression, the feature quantity of the physical quantity data is extracted. The decoding layer of the intermediate layer 54b is the layer for restoring the data (latent variable) whose dimension has been compressed by the encoding layer to the original dimension. Through this restoration, the physical quantity data representing the original features of the physical quantity data, that is, the features of the physical quantity that should be detected when the extruder 1 is normal, is restored. The output layer 54c is the layer for outputting the physical quantity data in which the features of the physical quantity are extracted in the encoding layer and the decoding layer and the normal features are reproduced.

[0082] As Figure 6 shown, the processing unit 51 causes the neural network of the first autoencoder 54 to perform machine learning so that the physical quantity data in the normal state input to the first autoencoder 54 is the same as the output physical quantity data. Specifically, various parameters such as the weight coefficients (coupling coefficients) between neurons constituting the neural network are optimized using the error backpropagation method, the steepest descent method, etc., so that the physical quantity data input in the normal state is the same as the reproduced physical quantity data.

[0083] When the physical quantity data in the normal state is input to the first autoencoder 54 that has been learned in this way, physical quantity data having substantially the same characteristics as the input physical quantity data is output. When the physical quantity data in the abnormal state is input to the first autoencoder 54, as Figure 6 shown, the physical quantity data in the normal state is output.

[0084] Figure 7 is a conceptual diagram showing the second autoencoder 55 in the detection stage, Figure 8 is a conceptual diagram showing the third autoencoder 56 in the detection stage.

[0085] As Figure 7 shown, the second autoencoder 55 is an autoencoder that reproduces and outputs the feature quantity that should be obtained from the normal extruder 1 when the feature quantity calculated based on a plurality of physical quantity data is input. Various combinations of a plurality of physical quantity data can be considered, but combinations of the power of the motor 13 for the extruder or the current and torque, and combinations of the torque and rotational speed of the motor 13 or the screw 11 are preferred.

[0086] As Figure 8 shown, the third autoencoder 56 is an autoencoder that reproduces and outputs the feature quantity that should be obtained from the normal extruder 1 when the feature quantity calculated based on the physical quantity data and the set value of the extruder 1 associated with the physical quantity data is input. Various combinations of the physical quantity data and the set value can be considered, but combinations of the raw material supply amount and the raw material supply amount set value, and combinations of the screw speed and the screw speed set value are preferred.

[0087] The structures of the second autoencoder 55 and the third autoencoder 56 are the same as that of the first autoencoder 54, so their detailed descriptions are omitted.

[0088] In addition, the first autoencoder 54, the second autoencoder 55, and the third autoencoder 56 may also be a CNN autoencoder having a CNN (Convolutional Neural Network), an RNN encoder having an RNN (Recurrent Neural Network), or an LSTM autoencoder having an LSTM (Long Short-Term Memory).

[0089] In addition, in the case of the CNN autoencoder, as the physical quantity data, image data representing a detection value waveform for a specified period including the current time can be input.

[0090] In addition, multiple first autoencoders 54, second autoencoders 55, and third autoencoders 56 can be prepared separately according to the type of the physical quantity data.

[0091] In addition, in the present embodiment, for ease of understanding the invention, the first autoencoder 54, the second autoencoder 55, and the third autoencoder 56 are separately described as different autoencoders, but if possible, they may also be constituted by one or two autoencoders.

[0092] Figure 9 It is a conceptual diagram showing an example of the record layout of the collection data DB57. The collection data DB57 includes a hard disk and a DBMS (Data Base Management System), and stores various physical quantity data collected from the extruder 1. For example, the collection data DB57 has columns of "No." (record number), "equipment ID", "operation date", "set value", and "physical quantity data".

[0093] The "equipment ID" column stores the equipment identifier of the extruder 1. The "operation date" column stores information indicating the year, month, date, etc. when various data are obtained and stored as records. The "set value" column stores the set value set for the extruder 1. The set value includes, for example, a raw material supply amount set value and a screw speed set value. The "physical quantity data" column stores time-series physical quantities indicating the operation state of the extruder 1, such as raw material supply amount, screw speed, extruder power, extruder torque, extruder current, resin pressure, resin temperature, etc.

[0094] <Machine learning process>

[0095] Figure 10This is a flowchart showing the learning process sequence of the autoencoder. The processing unit 51 of the state estimation device 5 acquires physical quantity data and set values related to the state of the extruder 1 from the data collection device 3 (step S11). The processing unit 51 stores the acquired physical quantity data in the collected data DB57. When the physical quantity data required for machine learning has been collected, the following processing is executed.

[0096] Next, the processing unit 51 performs machine learning processing of the first autoencoder 54 based on the acquired physical quantity data (step S12).

[0097] In addition, the processing unit 51 calculates feature quantities based on the acquired multiple physical quantity data (step S13), and performs machine learning processing of the second autoencoder 55 based on the calculated feature quantities (step S14).

[0098] Furthermore, the processing unit 51 calculates feature quantities based on the acquired physical quantity data and the set values associated with the physical quantity data (step S15), and performs machine learning processing of the third autoencoder 56 based on the calculated feature quantities (step S16). The calculation method of the feature quantities in step S16 is not particularly limited, and the physical quantity data can be scaled and transformed based on the set values.

[0099] After finishing the learning of the first to third autoencoders 54, 55, and 56, the processing unit 51 stores the data used to construct the learned first to third autoencoders 54, 55, and 56 in the storage unit 52 (step S17).

[0100] In addition, although the method of collecting physical quantity data to generate the first to third autoencoders 54, 55, and 56 has been described, it is also possible to perform machine learning of the first to third autoencoders 54, 55, and 56 while collecting physical quantity data during the operation of the extruder 1. Alternatively, it can be configured to perform additional learning of the first to third autoencoders 54, 55, and 56 at an appropriate timing based on the physical quantity data acquired during the operation of the extruder 1.

[0101] <State determination process>

[0102] Figure 11 This is a flowchart showing the processing sequence of the state estimation of the extruder 1 according to Embodiment 1. The processing unit 51 of the state estimation device 5 acquires physical quantity data and set values related to the state of the extruder 1 from the data collection device 3 (step S31).

[0103] Next, the processing unit 51 inputs the acquired physical quantity data into the first autoencoder 54 to reproduce the physical quantity data in the normal state (step S32). Then, the processing unit 51 compares the physical quantity data input to the first autoencoder 54 with the physical quantity data output from the first autoencoder 54, thereby determining the presence or absence of an abnormality in the extruder 1 (step S33).

[0104] In addition, the processing unit 51 may also calculate the degree of abnormality of the extruder 1 by calculating the difference between the physical quantity data input to the first autoencoder 54 and the physical quantity data output from the first autoencoder 54. The processing unit 51 may also use the difference between the value of the physical quantity data input to the first autoencoder 54 and the value of the physical quantity data output from the first autoencoder 54 at each moment in the time series as the residual, and calculate the cumulative value of the residual as the degree of abnormality of the extruder 1. In addition, the processing unit 51 may also calculate the sum of squared residuals and the root-sum square of the residuals based on the residuals as the degree of abnormality of the extruder 1.

[0105] Next, the processing unit 51 calculates a feature quantity based on the acquired multiple physical quantity data (step S34), inputs the calculated feature quantity into the second autoencoder 55, and thereby reproduces the feature quantity in the normal state (step S35). Then, the processing unit 51 determines the presence or absence of an abnormality in the extruder 1 by comparing the feature quantity input to the second autoencoder 55 with the feature quantity output from the second autoencoder 55 (step S36). In addition, the processing unit 51 may also calculate the degree of abnormality of the extruder 1 by calculating the difference between the feature quantity input to the second autoencoder 55 and the feature quantity output from the second autoencoder 55, in the same manner as in step S33.

[0106] Furthermore, the processing unit 51 calculates a feature quantity based on the acquired physical quantity data and the set value associated with the physical quantity data (step S37), inputs the calculated feature quantity into the third autoencoder 56, and thereby reproduces the feature quantity in the normal state (step S38). Then, the processing unit 51 determines the presence or absence of an abnormality in the extruder 1 by comparing the feature quantity input to the third autoencoder 56 with the feature quantity output from the third autoencoder 56 (step S39). In addition, the processing unit 51 may also calculate the degree of abnormality of the extruder 1 by calculating the difference between the feature quantity input to the third autoencoder 56 and the feature quantity output from the third autoencoder 56, in the same manner as in step S33.

[0107] Next, the processing unit 51 synthesizes the determination results of step S33, step S36, and step S39, calculates the presence or absence or degree of abnormality of the extruder 1 (step S40), and ends the processing. For example, when determining the presence or absence of an abnormality in each step, the processing unit 51 can determine the presence or absence of an abnormality in the extruder 1 based on the most determination results. When the degree of abnormality is calculated in each step, the processing unit 51 can calculate the average value, maximum value, etc. of the degree of abnormality as the comprehensively determined degree of abnormality.

[0108] As described above, according to the state estimation method and the like of the first embodiment, regardless of the differences in the model, operation method, and raw material of the extruder 1, the presence or absence or degree of abnormality of the extruder 1 can be determined.

[0109] Generally speaking, due to the differences in the model, operation method, and resin raw material of the extruder, the characteristics such as the range of the detected physical quantity data are different, so it cannot be dealt with in the abnormality detection using a threshold value. The state estimation method of the first embodiment is a structure that performs abnormality determination by reproducing the physical quantity data in the normal state using an autoencoder and comparing them. Therefore, regardless of the differences in the model, operation method, and resin raw material of the extruder, abnormalities can be detected.

[0110] In particular, in the first embodiment, by using the characteristic quantities calculated based on the torque and rotational speed of the motor 13 or the screw 11, the presence or absence or degree of abnormality of the extruder 1 can be calculated more accurately.

[0111] In addition, by using the characteristic quantities calculated based on the power or current and torque of the motor 13 for the extruder as multiple physical quantity data, the presence or absence or degree of abnormality of the extruder 1 can be calculated more accurately.

[0112] Moreover, in the first embodiment, by using the characteristic quantities calculated based on the raw material supply amount and the raw material supply amount set value, the presence or absence or degree of abnormality of the extruder 1 can be calculated more accurately.

[0113] In addition, by using the characteristic quantities calculated based on the screw speed and the screw speed set value, the presence or absence or degree of abnormality of the extruder 1 can be calculated more accurately.

[0114] Furthermore, in the first embodiment, an example of using an autoencoder to detect the abnormality of the extruder 1 is described, but it can also be configured to use a learning model, decision tree, random forest, SVM (Support Vector Machine), etc. algorithms constituted by other arbitrary neural networks, multi-layer perceptrons (MLP), etc. to determine the abnormality of the extruder 1.

[0115] In addition, in the present Embodiment 1, an example in which the state estimation device 5 as a server on the cloud executes state estimation processing has been described. However, it may also be configured such that the control device 15, the data collection device 3 such as a PLC, and the local computer connected to the data collection device 3 execute state estimation processing. In the case where the data collection device 3 executes state estimation processing, the communication unit 33 is not required.

[0116] Moreover, in the present Embodiment 1, an example of determining the presence or absence of an abnormality in the extruder 1 has been mainly described. However, it may also be configured to determine any state of the extruder 1. For example, it may be configured to determine an abnormal state of the extruder 1, a normal state of the extruder 1, an operating state of the extruder 1, and a state related to a combination thereof. Specifically, the autoencoder can be made to perform machine learning so that when physical quantity data or feature quantity is input, the physical quantity data or feature quantity that can be obtained from the extruder 1 in a specific abnormal state is reproduced and output. By comparing the obtained physical quantity data or feature quantity with the physical quantity data or feature quantity output from the autoencoder, it is possible to determine whether it is in a specific abnormal state. In addition, the autoencoder can be made to perform machine learning so that when physical quantity data or feature quantity is input, the physical quantity data or feature quantity that can be obtained from the extruder 1 in a specific operating state is reproduced and output. By comparing the obtained physical quantity data or feature quantity with the physical quantity data or feature quantity output from the autoencoder, it is possible to determine whether the extruder 1 is in a specific operating state.

[0117] (Embodiment 2)

[0118] In the state estimation system of Embodiment 2, the calculation method of the feature quantity based on a plurality of physical quantity data is different from that of Embodiment 1. The other configuration of the state estimation system is the same as that of the system of Embodiment 1. Therefore, the same reference numerals are given to the same parts, and detailed description thereof is omitted.

[0119] Figure 12 It is a conceptual diagram showing the feature quantity calculation model 58 and the second autoencoder 55 of Embodiment 2. The state estimation device 5 of Embodiment 2 stores the feature quantity calculation model 58 in the storage unit 52.

[0120] The feature quantity calculation model 58 has an input layer 58a, an intermediate layer 58b, and an output layer 58c. The feature quantity calculation model 58 is, for example, the same learning model as the first autoencoder 54.

[0121] For example, the processing unit 51 causes the neural network of the feature quantity calculation model 58 to perform machine learning so that the first physical quantity data in the normal state input to the feature quantity calculation model 58 is the same as the output first physical quantity data. In addition, the neural network of the feature quantity calculation model 58 is caused to perform machine learning so that the second physical quantity data in the normal state input to the feature quantity calculation model 58 is the same as the output second physical quantity data. The first physical quantity data is, for example, torque data acting on the screw 11. The second physical quantity data is, for example, the rotational speed of the screw 11.

[0122] The feature quantity calculation model 58 learned in this way is learned to be able to extract the features of each of the first physical quantity data and the second physical quantity data. As Figure 12 shown, when the first physical quantity data and the second physical quantity data are input to the feature quantity calculation model 58, the features of both the first physical quantity data and the second physical quantity data are extracted as latent variables. This latent variable corresponds to the feature quantity calculated based on the first physical quantity data and the second physical quantity data.

[0123] The processing unit 51 obtains a latent variable having the features of both the first physical quantity data and the second physical quantity data from the intermediate layer 58b of the feature quantity calculation model 58, uses the obtained latent variable as a feature quantity, and inputs it to the second autoencoder 55 in the same manner as in the first embodiment, thereby calculating the latent variable of the reproduction in the normal state. Then, the processing unit 51 determines the presence or absence and the degree of abnormality of the extruder 1 by comparing the latent variable input to the second autoencoder 55 with the latent variable output from the second autoencoder 55.

[0124] In addition, needless to say, the second autoencoder 55 is learned so that when a feature quantity as a latent variable is input, it outputs the latent variable that should be detected and calculated in the normal state.

[0125] As described above, according to the state estimation method and the like of the second embodiment, compared with the rule-based arithmetic processing, it is possible to calculate a more accurate feature quantity based on a plurality of physical quantity data, and thus it is possible to more accurately determine the state of the extruder 1.

[0126] In addition, in the second embodiment, torque and rotational speed are described as the first physical quantity data and the second physical quantity data, but the power or current of the motor 13 for the extruder may be used as the first physical quantity data, and torque may be used as the second physical quantity data.

[0127] (Embodiment 3)

[0128] The method for determining the state of the extruder 1 of the state estimation system according to the third embodiment is different from that of the first embodiment. The other structures of the state estimation system are the same as those of the system according to the first embodiment, so the same reference numerals are used for the same parts, and the detailed description is omitted.

[0129] Figure 13 It is a conceptual diagram showing the state estimation method of Embodiment 3. The state estimation device 5 of Embodiment 3 stores the autoencoders 354a and VAE 354b as the first autoencoder 54, and the CNN autoencoder 355a and the LSTM autoencoder 355b as the second autoencoder 55 in the storage unit 52.

[0130] In addition, the autoencoder 354a is an autoencoder that processes physical quantity data which is a numerical column as a time series. In contrast, the CNN autoencoder 355a is an autoencoder that processes physical quantity data represented by an image.

[0131] In addition, the processing unit 51 has a threshold determination unit 351a, a comparison processing unit 351b, and a comprehensive determination unit 351c as functional units.

[0132] The threshold determination unit 351a determines the presence or absence of an abnormality in the extruder 1 by comparing the physical quantity data acquired by the communication unit 53 with a prescribed threshold.

[0133] In addition, the threshold determination unit 351a may also compare the threshold with the value of the physical quantity data and calculate the cumulative value, sum of squares, and square root of the number of the parts exceeding the threshold as the degree of abnormality.

[0134] The comparison processing unit 351b determines the presence or absence of an abnormality in the extruder 1 or calculates the degree of abnormality by comparing the physical quantity data or feature quantity input to various autoencoders with the output physical quantity data or feature quantity. A machine learning model using a neural network such as an MLP may also be used to comprehensively determine the presence or absence of an abnormality or the degree of abnormality in the extruder 1. In addition, the determination processing based on the machine learning model and the determination processing based on rules may be combined for comprehensive determination.

[0135] The comprehensive determination unit 351c comprehensively determines the presence or absence of an abnormality in the extruder 1 based on the determination result of the threshold determination unit 351a and the determination result of the comparison processing unit 351b. For example, the presence or absence of an abnormality in the extruder 1 is determined according to the most determination results.

[0136] In addition, when calculating the degree of abnormality using the threshold determination unit 351a and the comparison processing unit 351b, the comprehensive determination unit 351c may also be configured to calculate the average value, maximum value, etc. of the degree of abnormality as the comprehensively determined degree of abnormality.

[0137] As described above, according to the state estimation method and the like of the present Embodiment 3, through the comprehensive determination processing based on the determination results obtained by using a variety of autoencoders, the state of the extruder 1 can be determined more accurately.

[0138] (Embodiment 4)

[0139] The state estimation system of Embodiment 4 is different from that of Embodiment 1 in that it discriminates the operating condition of the extruder 1 and uses an autoencoder corresponding to the operating condition to determine an abnormality of the extruder 1. Other configurations of the state estimation system are the same as those of the system of Embodiment 1, and thus the same reference numerals are assigned to the same parts and detailed descriptions are omitted.

[0140] Figure 14 FIG. is a block diagram showing a structural example of the state estimation device 5 of Embodiment 4. The state estimation device 5 of Embodiment 4 stores an operating condition discrimination model 59 for discriminating the operating condition of the extruder 1 in the storage unit 52. The operating condition corresponds to the operating process of the extruder, and the operating condition includes, for example, a stop state, a start-up operating state, a stable operating state, and a shutdown operating state.

[0141] The operating condition discrimination model 59 has an input layer, an intermediate layer, and an output layer (not shown). The operating condition discrimination model 59 is, for example, a learning model that outputs data representing the operating condition of the extruder 1 at the time when the physical quantity data is detected when the physical quantity data is input.

[0142] For example, learning data representing the operating condition is prepared, and the operating condition discrimination model 59 is made to perform machine learning using the learning data. The data representing the operating condition is, for example, data for unsupervised learning for a normal state, semi-supervised learning that extracts only abnormal states, or supervised learning with labels.

[0143] For example, in the case of supervised learning, learning data in which data representing the operating condition of the extruder 1 when the physical quantity data is obtained is attached to the physical quantity data as supervised data is prepared. Then, when the physical quantity data of the learning data is input to the operating condition discrimination model 59, the processing unit 51 makes the neural network of the operating condition discrimination model 59 perform machine learning so that the operating condition represented by the data output from the operating condition discrimination model 59 is consistent with the operating condition represented by the supervised data.

[0144] It is also possible to use physical quantity data obtained in a plurality of operating conditions such as a stop state, a start-up operating state, a stable operating state, and a shutdown operating state as learning data and make the operating condition discrimination model 59 perform unsupervised learning. For example, the operating condition discrimination model 59 is a model that clusters physical quantity data and outputs data representing the class (operating condition) to which the input physical quantity data belongs. In addition, it is also possible to use physical quantity data obtained in one operating condition as learning data and make the operating condition discrimination model 59 perform unsupervised learning.

[0145] Learning data including physical quantity data with tag data indicating the operating condition and physical quantity data without tag data may also be used to perform semi-supervised learning on the operating condition discrimination model 59. For example, physical quantity data with tag data indicating an abnormal state and physical quantity data without tag data for other states may also be used to perform semi-supervised learning on the operating condition discrimination model 59.

[0146] In addition, the storage unit 52 stores first to third autoencoders 54, 55, and 56 that have been learned for different operating conditions. For example, the storage unit 52 stores a first autoencoder 54 for determining an abnormality during startup operation, a first autoencoder 54 for determining an abnormality during stable operation, and a first autoencoder 54 for determining an abnormality during shutdown operation. The same applies to the second autoencoder 55 and the third autoencoder 56.

[0147] Figure 15 FIG. is a flowchart showing the processing sequence of state estimation of the extruder 1 according to Embodiment 4. The processing unit 51 of the state estimation device 5 acquires physical quantity data and set values related to the state of the extruder 1 from the data collection device 3 (step S431).

[0148] Next, the processing unit 51 uses the operating condition discrimination model 59 to discriminate the operating condition of the extruder 1 (step S432). Specifically, the processing unit 51 inputs the acquired physical quantity data into the operating condition discrimination model 59, and discriminates the current operating condition of the extruder 1 based on the data output from the operating condition discrimination model 59.

[0149] Then, the processing unit 51 selects the first to third autoencoders 54, 55, and 56 corresponding to the operating condition discriminated in step S432, and uses the selected autoencoders to determine the presence or absence of an abnormality in the extruder 1 or calculate the degree of abnormality (step S433). The method of abnormality determination using the first to third autoencoders 54, 55, and 56 is the same as in Embodiment 1. Then, the processing unit 51, in the same manner as in Embodiment 1, synthesizes the determination results obtained using each autoencoder, calculates the presence or absence of an abnormality or the degree of abnormality of the extruder 1 (step S434), and ends the processing.

[0150] As described above, according to the state estimation method and the like of the present Embodiment 4, an autoencoder corresponding to the operating condition of the extruder 1 can be selected to determine the presence or absence of an abnormality in the extruder 1 or calculate the degree of abnormality.

[0151] Means for solving the problems of the present disclosure are appended.

[0152] (Supplementary Note 1) A method for state estimation, which determines the state of an extruder by obtaining physical quantity data associated with the state of the extruder and inputting the obtained physical quantity data or a feature quantity based on the physical quantity data into a machine learning model. The machine learning model outputs information indicating the state of the extruder when the physical quantity data or the feature quantity based on the physical quantity data is input.

[0153] (Supplementary Note 2) According to the state estimation method described in Supplementary Note 1, the machine learning model is a model that outputs information indicating the state of the extruder when the physical quantity data and the set value of the extruder or data representing the static characteristics of the extruder are input, or when a feature quantity based on the physical quantity data and the set value of the extruder or data representing the static characteristics of the extruder is input. The state estimation method determines the state of the extruder by obtaining the physical quantity data associated with the state of the extruder and the set value of the extruder or data representing the static characteristics of the extruder, inputting the obtained physical quantity data and the set value of the extruder or data representing the static characteristics of the extruder into the machine learning model, or inputting a feature quantity based on the obtained physical quantity data and the set value of the extruder or data representing the static characteristics of the extruder into the machine learning model.

[0154] (Supplementary Note 3) According to the state estimation method described in Supplementary Note 1 or 2, the machine learning model includes an autoencoder. The autoencoder reproduces and outputs the physical quantity data or the feature quantity obtained from the extruder in a specified state when the physical quantity data or a feature quantity based on the physical quantity data is input. By inputting the obtained physical quantity data or feature quantity into the autoencoder and comparing the obtained physical quantity data or feature quantity with the physical quantity data or feature quantity output from the autoencoder, it is determined whether the extruder is in the specified state.

[0155] (Supplementary Note 4) According to the state estimation method described in any one of Supplementary Notes 1 to 3, the state of the extruder includes an abnormal state of the extruder, a normal state of the extruder, or an operating state of the extruder.

[0156] (Supplementary Note 5) According to the state estimation method described in any one of Supplementary Notes 1 to 4, the feature quantity is calculated based on a plurality of the physical quantity data.

[0157] (Supplementary Note 6) According to the state estimation method described in any one of Supplementary Notes 1 to 5, the extruder has a motor and a screw, and the feature quantity is calculated based on the rotational speed and torque of the motor or the screw.

[0158] (Supplementary Note 7) According to the state estimation method described in any one of Supplementary Notes 1 to 6, the extruder has a motor and a screw, and the characteristic quantity is calculated based on the power or current and torque of the motor.

[0159] (Supplementary Note 8) According to the state estimation method described in any one of Supplementary Notes 1 to 7, a set value of the extruder associated with the physical quantity data is obtained, and the characteristic quantity is calculated based on the physical quantity data and the set value associated with the physical quantity data.

[0160] (Supplementary Note 9) According to the state estimation method described in any one of Supplementary Notes 1 to 8, the extruder has a screw, and the associated physical quantity data and the set value include a combination of a raw material supply amount and a raw material supply amount set value, or a combination of a screw speed and a screw speed set value.

[0161] (Supplementary Note 10) According to the state estimation method described in any one of Supplementary Notes 1 to 9, multiple machine learning models are used to respectively determine the state of the extruder, and the presence or absence or degree of abnormality of the extruder is comprehensively determined by synthesizing each determination result.

Claims

1. A method for state estimation, characterized in that the state estimation method determines the state of the extruder by obtaining physical quantity data associated with the state of the extruder and inputting the obtained physical quantity data or a feature quantity based on the physical quantity data into a machine learning model, and the machine learning model outputs information indicating the state of the extruder when the physical quantity data or a feature quantity based on the physical quantity data is input.

2. The state estimation method according to claim 1, characterized in that the machine learning model is a model that outputs information indicating the state of the extruder when the physical quantity data, the set value of the extruder or data representing the static characteristics of the extruder is input, or when a feature quantity based on the physical quantity data, the set value of the extruder or data representing the static characteristics of the extruder is input. The state estimation method determines the state of the extruder by obtaining the physical quantity data associated with the state of the extruder and the set value of the extruder or data representing the static characteristics of the extruder, and inputting the obtained physical quantity data and the set value of the extruder or data representing the static characteristics of the extruder into the machine learning model, or inputting a feature quantity based on the obtained physical quantity data and the set value of the extruder or data representing the static characteristics of the extruder into the machine learning model.

3. The state estimation method according to claim 1, characterized in that the machine learning model includes an autoencoder, and the autoencoder reproduces and outputs the physical quantity data or the feature quantity obtained from the extruder in a specified state when the physical quantity data or a feature quantity based on the physical quantity data is input. The state estimation method determines whether the extruder is in the specified state by inputting the obtained physical quantity data or feature quantity into the autoencoder and comparing the obtained physical quantity data or feature quantity with the physical quantity data or feature quantity output from the autoencoder.

4. The state estimation method according to claim 1, characterized in that the state of the extruder includes an abnormal state of the extruder, a normal state of the extruder or an operating state of the extruder.

5. The state estimation method according to claim 1, characterized in that the feature quantity is calculated based on a plurality of the physical quantity data.

6. The state estimation method according to claim 5, characterized in that the extruder has a motor and a screw, the feature quantity is calculated based on the rotational speed and torque of the motor or the screw.

7. The state estimation method according to claim 5, characterized in that the extruder has a motor and a screw, the feature quantity is calculated based on the power or current and torque of the motor.

8. The state estimation method according to claim 1, characterized in that the set value of the extruder associated with the physical quantity data is obtained. The characteristic quantity is calculated based on the physical quantity data and the set value associated with the physical quantity data.

9. The state estimation method according to claim 8, wherein the extruder has a screw, the associated physical quantity data and the set value include a combination of a raw material supply amount and a raw material supply amount set value, or a combination of a screw speed and a screw speed set value.

10. The state estimation method according to any one of claims 1 to 9, wherein a plurality of the machine learning models are used to respectively determine the state of the extruder, and the respective determination results are integrated to comprehensively determine the presence or absence or degree of abnormality of the extruder.

11. A state estimation device, wherein the state estimation device includes a processing unit, the processing unit performs the following processing: by acquiring physical quantity data associated with the state of the extruder and inputting the acquired physical quantity data or a characteristic quantity based on the physical quantity data to a machine learning model, the state of the extruder is determined, the machine learning model outputs information indicating the state of the extruder when the physical quantity data or a characteristic quantity based on the physical quantity data is input.

12. A program product, wherein the program product includes a computer program for causing a computer to perform the following processing: by acquiring physical quantity data associated with the state of the extruder and inputting the acquired physical quantity data or a characteristic quantity based on the physical quantity data to a machine learning model, the state of the extruder is determined, and the machine learning model outputs information indicating the state of the extruder when the physical quantity data or a characteristic quantity based on the physical quantity data is input.

13. A machine learning method, wherein physical quantity data associated with the state of the extruder is acquired, and a machine learning model is generated based on the acquired physical quantity data, and the machine learning model outputs information indicating the state of the extruder when the physical quantity data or a characteristic quantity based on the physical quantity data is input.

Images