Information processing method, information processing device, and non-transitory computer-readable recording medium

Abstract

The present invention discloses an information processing method, an information processing device, and a non-transitory computer-readable recording medium. The information processing device includes a controller. The controller is configured to obtain a measurement value of a sensor provided in a mechanical device. The controller is configured to generate a first model by machine learning using the measurement value of the sensor measured in a first time period of the mechanical device, and store the first model in a storage unit. The controller is configured to generate a second model by machine learning using the measurement value of the sensor measured in a second time period after a trigger event has occurred in the mechanical device, and store the second model in the storage unit. The controller is configured to determine the state of the mechanical device by using the measurement value of the sensor measured in an evaluation period and the first model and the second model stored in the storage unit.

Description

Technical Field

[0001] The present invention relates to an information processing method, an information processing device, etc. for detecting a sign of a failure of a mechanical device. Background Art

[0002] The operating state of a mechanical device may change every second depending on changes in the states of its components, etc. If a state in which the operating state is within an acceptable range is called a normal state, and a state in which the operating state is outside the acceptable range is called an abnormal state in consideration of the purpose of use of the mechanical device, then, for example, in the case of a manufacturing machine, failures such as the manufacture of defective products and the stoppage of a manufacturing line may occur in the abnormal state.

[0003] For example, in the field of machine tools, attempts have been made to detect abnormal conditions using a detection method using a learning model. However, in the case of a machine tool, the operation mode of the motor during machining, the type of tool used for machining, the type of workpiece to be machined, etc., change depending on the type of machining, and therefore it is difficult to create a learning model that has learned all types of machining for general use.

[0004] Therefore, Japanese Patent Laid-Open No. 2019-67136 discloses a method: the method prepares multiple learning models corresponding to different types of processing, and selects one of the multiple learning models according to the operating conditions and environmental conditions of the machine tool to perform abnormality detection corresponding to the type of processing (in other words, the operating state).

[0005] In addition, Japanese Patent Laid-Open No. 2019-28565 discloses a technology that generates a model of a normal state by machine learning using data of a period when a mechanical device is in a normal state and evaluates the deviation between the operating state and the normal state model to detect signs of a fault. In addition, it is disclosed that in the case of assuming an error detection, that is, in the case where the mechanical device is in a normal state although the deviation of the evaluation is large, the model is updated by adding data of a period corresponding to the error detection to the learned data of the normal period and performing additional machine learning.

[0006] As described above, in machine tools and the like, since failures such as the manufacture of defective products and the stoppage of a manufacturing line may occur in an abnormal state, it is important to detect the occurrence of the abnormal state quickly. In view of this, the abnormality detection method of Japanese Patent Laid-Open No. 2019-67136 is expected to be effective to some extent. However, it is more important to prevent the abnormal state from occurring.

[0007] In machine tools and the like, in order to suppress the occurrence of abnormal conditions as much as possible, maintenance operations are generally performed regularly or irregularly even when the same operation is repeatedly performed. Although shortening the intervals between maintenance operations is effective for increasing preventive safety, since the manufacturing machine and the like are stopped during the maintenance operations, when the frequency of maintenance operations excessively increases, the operating rate of the manufacturing machine and the like decreases. Therefore, it is preferable that such a state can be detected when the machine and the like are still in a normal state but the occurrence of an abnormal state is approaching. This is because, if the approach of the occurrence of an abnormal state can be detected, that is, if the occurrence of a failure can be predicted, then the maintenance operation of the machine and the like can be performed at that time, so that an excessive decrease in the operating rate can be suppressed.

[0008] In view of this, the method of detecting abnormality disclosed in Japanese Patent Laid-Open No. 2019-67136 is not necessarily a method that enables effective detection of the approach of the occurrence of an abnormal state (i.e., effective detection of signs of a fault).

[0009] Meanwhile, in the method for detecting a sign of a fault disclosed in Japanese Patent Laid-Open No. 2019-28565, a model of a normal state is generated in advance, and the model is updated by additional learning. However, it is actually difficult to achieve high detection accuracy by applying the method of Japanese Patent Laid-Open No. 2019-28565 to the prediction of faults in manufacturing machines, etc.

[0010] First, a state change occurring in a maintenance operation of a manufacturing machine, etc. will be described. Here, for example, in the case where the same processing operation is continuously and repeatedly performed by a manufacturing machine, etc., the state of the manufacturing machine, etc. at the start of the processing operation will be referred to as an initial state. That is, the initial state is one mode of the normal state of the manufacturing machine, etc.

[0011] As described above, in a manufacturing machine or the like, even in the case where the same operation is continuously and repeatedly performed, a maintenance operation is performed regularly or irregularly, but when the maintenance operation is performed, the state of the machine changes. It should be noted here that when a maintenance operation is performed in a manufacturing machine or the like, the machine may return to a state extremely close to the initial state, or may change to a state different from the initial state within the range of a normal state.

[0012] For example, a manufacturing machine or the like may include a part that needs to be periodically cleaned to maintain predetermined performance, but since the cleaning operation takes time, a plurality of the same kind of parts are prepared and used alternately in some cases. That is, by cleaning a used part detached from the machine while using another part attached to the machine and by replacing only the part during the maintenance operation, the time required for the maintenance operation, that is, the time during which the manufacturing machine is stopped, is shortened.

[0013] In this case, even the same kind of parts may have variations within the tolerance range, therefore, when the parts are replaced in the maintenance operation, the machine may be restored to a state extremely close to the initial state, or may be changed to a state different from the initial state. In addition, even if the parts are not replaced in the maintenance operation and the exact same parts are cleaned and reattached to the machine, the state of the parts is not necessarily completely restored to the original state, and the state of the machine after the maintenance operation may be different from the initial state. In addition, similarly, even in the case where the parts are replaced with unused brand-new parts in each maintenance operation, the state of the machine after the maintenance operation may be a state extremely close to the initial state or a state different from the initial state, depending on the variations of the brand-new parts.

[0014] Furthermore, since a manufacturing machine or the like is composed of a huge number of parts, even if one part is refreshed to the same state as the initial state during maintenance operation, the states of other parts have changed according to the usage history. Therefore, the balance of the machine as a whole may not return to the initial state.

[0015] In addition, in manufacturing machines, etc., maintenance operations are not necessarily the same every time. For example, there are cases where only component A is replaced among a large number of components such as parts, and there are cases where component A and component B are replaced at the same time if necessary.

[0016] As described above, each time maintenance operation is performed, the balance between the states of a large number of components constituting the machine or their usage history changes, and therefore the state of the machine immediately after maintenance may vary in various ways within the range of a normal state.

[0017] Therefore, although a learning model of an initial state as one mode of a normal state can be generated by causing a learning machine to learn data immediately after starting use, it is not easy to generate in advance a learning model that covers all changes in the normal state caused by maintenance operations.

[0018] In addition, even if a learning model based on the data of the initial state is generated as disclosed in Japanese Patent Laid-Open No. 2019-28565 and then the learning model is updated by adding data of the time period corresponding to the error detection to the data of the initial state to perform additional learning, the accuracy of determining signs of abnormal conditions does not necessarily increase significantly.

[0019] For example, there may be a case where the state of a component is changed by a first maintenance operation, so that the value of a detection parameter related to the component becomes discontinuous, and according to the learning model that has learned the initial state, the state immediately after the maintenance is determined to be a sign of a fault. In this case, this is an erroneous determination, and therefore, in the method of Japanese Patent Laid-Open No. 2019-28565, the learning model is updated by performing additional learning by adding data corresponding to the error detection to the data of the initial state. However, the learning model after the update may be a model that excessively reduces the deviation in the period in which the error detection has occurred. In order to increase the detection accuracy of the symptom, it is necessary to adjust the parameters according to the complexity of the data input when performing machine learning, and if additional learning is performed without adjusting the parameters of the machine learning, the detection accuracy may be reduced.

[0020] However, checking and adjusting parameters by machine learning experts and updating the learning model every time an error detection occurs is a huge burden on the management of manufacturing machines and the like including machine learning, and is not practical.

[0021] In addition, if a learning model is newly generated by using only data for the period corresponding to error detection without performing additional learning, there is a possibility that error detection occurs according to the newly generated learning model when the state of the machine returns to a state extremely close to the initial state after the next maintenance operation.

[0022] Therefore, for systems (such as manufacturing machines, etc.) that require maintenance operations for use and in which the state of detection parameters can be changed by maintenance operations in various ways within the range of normal states, it has been desired to enable an information processing method and an information processing device to detect signs of the occurrence of a fault with high accuracy. Summary of the invention

[0023] According to a first aspect of the present invention, an information processing method for a controller to determine a state of a mechanical device, the information processing method comprising: a measuring step, in which the controller obtains a measurement value of a sensor provided in the mechanical device; a first storage step, in which the controller generates a first model by machine learning using the measurement value of the sensor measured in a first time period, and stores the first model in a storage unit; a second storage step, in which the controller generates a second model by machine learning using the measurement value of the sensor measured in a second time period after a trigger event has occurred in the mechanical device, and stores the second model in the storage unit; and a determination step, in which the controller determines the state of the mechanical device by using the measurement value of the sensor measured in an evaluation period and the first model and the second model stored in the storage unit.

[0024] According to a second aspect of the present invention, an information processing device includes a controller. The controller is configured to: obtain a measurement value of a sensor provided in a mechanical device; generate a first model by machine learning using the measurement value of the sensor measured in a first period of the mechanical device, and store the first model in a storage unit; generate a second model by machine learning using the measurement value of the sensor measured in a second period after a trigger event has occurred in the mechanical device, and store the second model in the storage unit; and determine the state of the mechanical device by using the measurement value of the sensor measured in an evaluation period and the first model and the second model stored in the storage unit.

[0025] Further features of the present invention will become apparent from the following description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a functional block diagram illustrating functional blocks included in a failure prediction apparatus according to an exemplary embodiment.

[0027] Figure 2 is a graph used to describe the extraction of feature values.

[0028] Figure 3A is a diagram illustrating an example of learning data according to an exemplary embodiment.

[0029] Figure 3B is a diagram illustrating an example of machine learning according to an exemplary embodiment.

[0030] Figure 4A is a diagram illustrating an example of evaluation data according to an exemplary embodiment.

[0031] Figure 4B is a diagram illustrating an example of calculation of a degree of deviation according to an exemplary embodiment.

[0032] Figure 5 is a diagram showing an example of a graph representing a degree of deviation plotted against time.

[0033] Figure 6 is a diagram for describing a mode of detecting a sign of a failure by using a first learning model and a second learning model in combination.

[0034] Figure 7 is a flowchart showing the order of steps of a fault prediction method according to an exemplary embodiment.

[0035] Figure 8 is an example of an image displayed in the failure prediction apparatus according to the exemplary embodiment.

[0036] Fig. 9A An example of a screen for managing settings of a learning model in the failure prediction apparatus according to the exemplary embodiment is shown.

[0037] Fig. 9B A screen for displaying a list of post-learning models for managing the post-learning models in a fault prediction apparatus according to an exemplary embodiment is shown.

[0038] Fig. 10A is a diagram showing a simplified structure of an injection molding machine serving as an example of a target machine.

[0039] Fig. 10B : is a diagram showing an example of replacement of parts in a maintenance operation of an injection molding machine serving as an example of a target machine.

[0040] Fig.11 is a diagram for describing a process for determining the presence / absence of a symptom of a failure.

[0041] Fig.12 is a schematic diagram showing an example of a hardware configuration of a failure prediction device according to an exemplary embodiment. DETAILED DESCRIPTION

[0042] An information processing method serving as a control method of a manufacturing machine or the like and an information processing apparatus serving as a control apparatus serving as exemplary embodiments of the present invention will be described with reference to the drawings.

[0043] It should be noted that the "manufacturing machine, etc." used here does not necessarily refer only to machines, equipment, manufacturing systems, etc. related to the manufacture of products. The term refers to machines, equipment, systems, etc. that require maintenance operations for use and various parameters indicating their status can be changed by maintenance operations within the range of normal status. In the description of the exemplary embodiments below, for convenience, the control method and the control device may be referred to as a fault prediction method, a fault prediction determination, a fault prediction device, etc.

[0044] It is to be noted that, unless otherwise described, elements denoted by the same reference numerals in the drawings referred to in the following description have the same functions.

[0045] Exemplary Embodiments

[0046] Configuration of fault prediction device

[0047] Figure 1 is a schematic functional block diagram showing functional blocks included in the fault prediction device according to the present exemplary embodiment. Figure 1 Components required for describing the features of the present exemplary embodiment are shown as functional blocks, and illustration of general components that are not directly related to the principle of problem solving of the present invention is omitted. In addition, Figure 1 Each component shown in the figure is conceptual in function and is not necessarily physically configured as shown. For example, the specific mode of segmentation and integration of functional blocks is not limited to the examples shown, and all or part of them can be functionally or physically segmented or integrated into any unit according to various loads, usage conditions, etc.

[0048] The fault prediction device 2 includes a controller 20 and a storage unit 30, and each functional block is composed of software or some hardware of a computer. For example, among the functional blocks included in the controller 20, the collection unit 201 to the post-learning model adding unit 208 are implemented by executing a computer program stored in an auxiliary storage device and loaded into a main storage device by a processor such as a central processing unit (CPU) or a graphics processing unit (GPU). In addition, the storage unit 30 is implemented by a storage device such as a hard disk drive or a non-volatile memory, and stores the collected data 301, the feature value data 302, the learning data 303, the evaluation data 304, the learning model, the deviation 306, the additional learning data 307, etc.

[0049] Fig.12 An example of the hardware configuration of the fault prediction device according to the present exemplary embodiment is schematically shown. Fig.12As shown in , the fault prediction device may include computer hardware including a CPU 1601 used as a main controller, a read-only memory (ROM) 1602 used as a storage device, and a random access memory (RAM) 1603. ROM 1602 can store information such as a learning program and an inference algorithm for implementing a fault prediction method to be described later. In addition, when executing the control process of the fault prediction method, RAM 1603 is used as a work area of ​​CPU 1601, etc. In addition, an external storage device 1606 is connected to the control system. The external storage device 1606 is composed of a hard disk drive (HDD), a solid-state device (SSD), or an external storage device of another system installed on the network.

[0050] The processing program of the CPU 1601 for implementing the control process of the present exemplary embodiment to be described later can be stored in a storage unit such as an external storage device 1606 constituted by an HDD, an SSD, etc. or an electrically erasable programmable ROM (EEPROM) area such as a ROM 1602. In this case, the processing program of the CPU 1601 for implementing the control process to be described later can be supplied to each of the above-mentioned storage units via a network interface (NIF) 1607, and can be updated to a new or different program. Alternatively, the processing program of the CPU 1601 for implementing the control process to be described later can be supplied to each of the above-mentioned storage units via storage media such as various magnetic disks, optical disks, and flash memories and their drive devices, and their contents can be updated. Various storage media, storage units, and storage devices storing the control program of the CPU 1601 for implementing the above-mentioned control process each constitute a computer-readable recording medium storing the control process of the present invention.

[0051] CPU 1601 is connected to Figure 1 Although the sensor group 10 is shown in FIG. Fig.12 The sensor group 10 is shown as being directly connected to the CPU 1601, but the sensor group 10 may be connected to the CPU 1601 via, for example, a so-called general purpose interface bus (GPIB, such as IEEE488). Furthermore, the sensor group 10 may be connected to the CPU 1601 via a network interface 1607 and a network 1608.

[0052] The network interface 1607 may be configured by, for example, using a communication standard for wired communication such as IEEE 802.3 or a communication standard for wireless communication such as IEEE 802.11 or IEEE 802.15. The CPU 1601 may communicate with other devices 1104 and 1121 via the network interface 1607. The devices 1104 and 1121 may be, for example, an overall control device such as a programmable logic controller (PLC) or a sequencer or a management server provided in a manufacturing line for manufacturing control or management.

[0053] exist Fig.12 In the example shown in Figure 1 The operation unit 1604 and the display device 1605 corresponding to the input device 3 and the display device 4 shown in FIG1 are connected to the CPU 1601 as a user interface device (UI device). The operation unit 1604 may be composed of a terminal (such as a handheld terminal), a device (such as a keyboard, a jog dial, a mouse, a pointing device, etc.), or a control terminal including at least one of these. The display device 1605 may be any display as long as the display device 1605 can display information related to the learning model, evaluation data, etc. to be described later on the screen, and, for example, a liquid crystal display device may be used.

[0054] Return to Reference Figure 1 , a mechanical device 1 as a mechanical device serving as an object of fault prediction determination is communicatively connected to the fault prediction apparatus 2. Examples of the mechanical device 1 include an injection molding machine that heats and melts a plastic material, injects the melted plastic material into a mold cavity, and molds the injected material, but the mechanical device 1 is not limited thereto.

[0055] Fig. 10A A simplified structure of an injection molding machine used as an example of a mechanical device 1 is shown. An injection molding machine is a machine that, for example, heats and melts a pellet of plastic material, pressurizes the molten plastic material by a screw, and molds the plastic material by injecting the plastic material into a cavity defined by a mold. An injection molding machine is a complex device consisting of many components, and if the injection molding machine is used for a long period of time, for example, the properties of the molten material may change and the molten material may adhere to components such as screws. If adhesion develops, an abnormal state, i.e., a failure, may eventually occur in which an appropriate amount of material cannot be injected into the cavity and a defective product is manufactured. There are other factors that can cause abnormal states, and maintenance operations (such as disassembling and cleaning the machine or replacing parts) take a long time. Therefore, the fault prediction device of the present exemplary embodiment detects signs and issues a notification before the injection molding machine reaches an abnormal state (i.e., a failure), so that the user can perform maintenance operations at appropriate timing.

[0056] Return to reference again Figure 1 , the fault prediction device 2 is configured to collect data from a sensor group 10 provided in the mechanical device 1, and collect operation information 11 from a controller of the mechanical device 1. The data collected from the sensor group 10 are measurement data related to various parameters indicating the state of the mechanical device 1, such as the temperature of each part of the injection molding machine, the injection pressure of the molten plastic material, the load, position, vibration and sound of the actuator, etc. The load of the actuator is represented by a current value.

[0057] Function block operation

[0058] Next, functional blocks included in the failure prediction device 2 will be described sequentially.

[0059] The collecting section 201 collects sensor data from the sensor group 10 of the mechanical device 1 at a predetermined cycle, and stores the collected data in the storage section 30 as collected data 301. Measurement data related to parameters that can be used as determining factors for detecting whether the occurrence of an abnormality (i.e., a failure) in the mechanical device 1 is approaching is collected.

[0060] The feature value generating unit 202 serving as a feature value extracting unit extracts or calculates a feature value representing a feature of the operating state of the mechanical device 1 by processing the collected data 301, and stores the feature value in the storage unit 30 as feature value data 302. For example, as feature value data, the maximum value and / or minimum value of the measurement value of the sensor collected in one operating cycle of the mechanical device 1 can be extracted, or the average value thereof can be calculated. Alternatively, for example, the feature value data can be obtained by performing an integral conversion of the value of the sensor in a predetermined period to a time series frequency region. In addition, the feature value data can be a differential value or a secondary differential value of the sensor value arranged in a time series with respect to time. In addition, in the case where the measurement value of the sensor itself (i.e., the raw data) is sufficient as a determining factor for detecting whether the occurrence of an abnormality is approaching, these measurement values ​​themselves can be used as feature value data 302. In the present exemplary embodiment, as Figure 2 As shown in , a mechanical device 1 serving as an object of fault prediction includes a sensor group 10 including sensors 1 to 3 for measuring various parameters for detecting the state of the machine. In the fault prediction method according to the present exemplary embodiment, feature values ​​such as periodicity and correlation are extracted from sensor data obtained from the sensor group 10. For example, the feature value generation section 202 extracts or calculates feature values ​​based on the measurement data of sensors 1 to 3, thereby generating time series feature values ​​1 to 3 corresponding to the sampling timing of the sensors, and stores the feature values ​​1 to 3 as feature value data 302 in the storage section 30.

[0061] The extraction unit 203 extracts data of an arbitrary time period from the feature value data 302 stored in the storage unit 30. In the present exemplary embodiment, a data set of a "normal time period" is extracted from the feature value data 302, and the extracted data set is stored in the storage unit 30 as the learning data 303. Here, the "normal time period" refers to a predetermined time period set by selecting a time period in which the mechanical device 1 is clearly in a normal state (such as an initial stage after the start of use of the mechanical device 1 or a period immediately after maintenance). That is, the normal time period is a predetermined time period in which the occurrence of a fault is not yet approaching. In the present exemplary embodiment, the "normal time period" is the first three days of the period in which the mechanical device 1 is in a normal state. Here, the learning data 303 is a data set of the feature value data 302 generated from the collected data of the first three days after the start of use of the mechanical device 1 (i.e., from the collected data of the initial state).

[0062] Furthermore, when determining whether the occurrence of abnormality is approaching, the extraction section 203 extracts a data set of a period as an evaluation target from the feature value data 302 , and stores the extracted data set in the storage section 30 as evaluation data 304 .

[0063] Furthermore, when the extraction section 203 has received an instruction to extract additional data from the determination section 206 , the extraction section 203 extracts a data set of an additional period from the feature value data 302 , and stores the extracted data set as additional learning data 307 in the storage section 30 .

[0064] The learned model generation unit 204 performs machine learning based on the learning data 303 which is a data set of feature value data 302 of the initial state of the mechanical device 1 after starting to use, and generates a learned model 1 as a first learned model. Then, the learned model generation unit 204 stores the learned model 1 in a learned model storage area 305 serving as a learned model storage unit.

[0065] The calculation unit 205 calculates the degree of deviation 306 from the normal state of the mechanical device 1 based on the evaluation data 304 and the output obtained by inputting the evaluation data 304 into the learned model stored in the learned model storage area 305. The calculation unit 205 is implemented by applying the learned model stored in the learned model storage area 305 to a platform capable of performing inference processing through machine learning. The processing algorithm of the calculation unit 205 can be, for example, inference processing using a multi-layer neural network, or a learning algorithm known as machine learning (such as a support vector machine or a Gaussian mixture model). In the case where a plurality of learned models are stored in the learned model storage area 305, the calculation unit 205 calculates the degree of deviation separately for each learned model.

[0066] The determination section 206 refers to the deviation 306 calculated by the calculation section 205 using each learned model, and when all calculated deviations are equal to or greater than the determination threshold, determines that there is a sign of a failure, and issues a notification request instruction to the notification section 207 .

[0067] It is to be noted that, in the case where it is necessary to additionally generate a new post-learning model, the determination section 206 instructs the extraction section 203 to extract feature value data of an additional period as the additional learning data 307 .

[0068] In the case where the notification unit 207 has received a notification request instruction based on the determination that there is a sign of a fault from the determination unit 206, the notification unit 207 notifies the external device (such as the display unit) of the alarm. As a method of notification, for example, the notification can be performed by lighting a lamp of the mechanical equipment, generating a warning sound, displaying a notification message on a display, or sending an e-mail. In addition, the notification can be displayed on the display device 4 via the display unit 210.

[0069] In the case where the additional learning data 307 has been extracted, the learned model adding unit 208 generates a learned model 2 by machine learning based only on the additional learning data 307, and adds the learned model 2 to the learned model storage area 305. That is, similar to the learned model generating unit 204 that has generated the learned model 1, the learned model adding unit 208 generates the learned model 2. Thereafter, each time the additional learning data 307 is extracted, the learned model adding unit 208 generates a new learned model by performing machine learning based on the additional learning data, and adds the learned model 3, the learned model 4, ... to the learned model storage area 305.

[0070] The input unit 209 receives user instructions input through the input device 3 (such as a keyboard or a mouse), for example, a user instruction to select feature value data 302 to be used to generate a learning model, or a user instruction to specify a time period from which learning data 303 is to be extracted.

[0071] The display device 4 included in the failure prediction apparatus 2 is constituted by a display device such as a liquid crystal display as described above. Note that the display device 4 may be an external device connected to the failure prediction apparatus 2.

[0072] The display section 210 processes data such as the collected data 301 to the additional learning data 307 stored in the storage section 30 to make the data visually understandable to the user, and causes the display device 4 such as a liquid crystal display to display the processed data. For example, the display section 210 may process the data into a form of a graph in which the data is plotted with respect to time or into a form of a table, and causes the display device 4 to display the processed data.

[0073] Fault prediction methods

[0074] Next, we will refer to Figure 7 The flowcharts describe the operation of the failure prediction device and the failure prediction method according to the present exemplary embodiment.

[0075] After starting use, first, in step S101 , the controller 20 determines whether this is the first use of the failure prediction device 2 for the operation to be started on the machine 1 based on the operation information 11 input from the machine 1 .

[0076] Here, if the use is the first time, that is, if the result of step S101 is "yes", the controller 20 sets a "normal period" in step S102. In step S102, a specific period such as "three days from date xx / xx" is set as the first normal period for the first use.

[0077] Next, in step S103, the extraction section 203 extracts feature value data of a normal period from the feature value data 302 as learning data 303. Then, the learned model generation section 204 generates a learned model 1 as a first learned model by a machine learning algorithm using the learning data 303.

[0078] Here, temporarily stop Figure 7 A description of the flowchart will be given, and a method for generating a learned model will be described.

[0079] Generation of the model after learning

[0080] The machine learning of this exemplary embodiment will be described by taking a three-layer neural network as an example. In this exemplary embodiment, Figure 2 The three continuous time series data of eigenvalues ​​1 to 3 shown in are used to generate a learned model that estimates and outputs the eigenvalues ​​1 to 3 at the next timing.

[0081] First, the data structure of the learning data 303 used for machine learning will be described. The extraction section 203 extracts feature value data of a normal period as the learning data 303, and the learning data 303 has a data structure including a plurality of pieces of learning data.

[0082] First, prepare for normal hours. Figure 2 The time series feature value data corresponding to the sensor sampling timing shown on the right side of , that is, feature values ​​1 to 3. Based on this time series data, a plurality of pieces of learning data in which an input signal and teaching data are combined are generated.

[0083] Figure 3AA piece of learning data is shown as an example. Continuous eigenvalue data corresponding to three sampling times (i.e., eigenvalues ​​1 to 3) are extracted from the time series eigenvalue data and prepared as input signals X1 to X9 of the neural network. In addition, T1, T2, and T3, which are eigenvalue data corresponding to subsequent sampling timings of X3, X6, and X9, are extracted and prepared as teaching data. A set of input signals X1 to X9 and teaching data T1 to T3 are set as one piece of learning data.

[0084] A plurality of pieces of learning data based on the feature values ​​of the entire “normal period” are prepared by generating each piece of learning data while changing the sampling timing of the input signal and the teaching data by an increment of one sampling time.

[0085] Figure 3B An example of machine learning using one piece of learning data is shown. The neural network learns the weight of the internal function so as to minimize the difference between the output data Y and the teaching data T. As multiple pieces of prepared learning data are sequentially added to the neural network one by one, the neural network updates the weight of the internal function so as to minimize the difference between the output data Y and the teaching data T. By completing the update of the weight using all the learning data corresponding to the "normal period", a post-learning model is generated in the neural network.

[0086] As described above, the term "normal period" refers to a period in which a normal state is expected to continue for a long time, such as the initial stage after starting to use the mechanical equipment 1 or the period immediately after a maintenance operation. Therefore, the learned model generated here is the following learned model: the learned model is capable of predicting and outputting the fourth eigenvalue after three eigenvalue data consecutive in time series in the period in which the normal state is expected to continue for a long time when input.

[0087] Return to Reference Figure 7 In step S104 , the learned model 1 generated in step S103 is stored in the learned model storage area 305 .

[0088] When the storage of the learned model 1 is completed in step S104, the processing is transferred to step S105. It should be noted that in the case where it has been determined in step S101 that this is not the first time the fault prediction device 2 is used (that is, in the case where the result of step S101 is "No"), the generation of the learned model 1 has been completed, so steps S102 to S104 are skipped and the processing is transferred to step S105.

[0089] In step S105, the calculation section 205 calculates the deviation 306, which is an index for determining whether the occurrence of a failure of the mechanical device 1 is close in the evaluation period as the time point at which the state of the mechanical device 1 is to be evaluated. The deviation 306 can also be referred to as an index indicating how much the state of the mechanical device 1 in the evaluation period is different from the initial stage of the normal state (i.e., in the normal period).

[0090] Calculation of Deviation

[0091] The extraction unit 203 extracts feature value data corresponding to a time point (ie, evaluation period) when the state of the machine 1 is desired to be evaluated from the feature value data 302 as evaluation data 304. For example, when the current state is desired to be evaluated, feature value data of the most recent past is extracted.

[0092] Figure 4A As an example, one piece of evaluation data is shown. Continuous eigenvalue data corresponding to three sampling times (i.e., eigenvalues ​​1 to 3) are extracted from the time series eigenvalue data and prepared as input signals X1 to X9 of the model after learning. In addition, eigenvalue data y1, y2, and y3 corresponding to the sampling timings subsequent to the sampling timings of X3, X6, and X9 are extracted from the time series eigenvalue data and prepared as measurement values. A set of input signals X1 to X9 and measurement values ​​y1 to y3 are set as one piece of evaluation data. Although in Figure 4A Only one piece of evaluation data including X1 to y1 corresponding to eigenvalue 1, X4 to y2 corresponding to eigenvalue 2, and X7 to y3 corresponding to eigenvalue 3 is shown, but it is preferable to prepare a plurality of evaluation data by using input signals and measurement values ​​of different sampling timings.

[0093] like Figure 4B As shown in , the fault prediction device 2 inputs the input signals X1 to X9 included in the evaluation data into the learned model stored in the learned model storage area 305, and obtains the output signals Y1, Y2 and Y3 from the learned model. That is, X1 to X3 of eigenvalue 1, X4 to X6 of eigenvalue 2, and X7 to X9 of eigenvalue 3 are input into the learned model, and the output signal Y1 corresponding to the eigenvalue 1, the output signal Y2 corresponding to the eigenvalue 2, and the output signal Y3 corresponding to the eigenvalue 3 are output as regression values ​​(i.e., estimated values). This step is performed on all the learned models stored in the learned model storage area 305, and the output signal is obtained from each learned model.

[0094] The calculation section 205 calculates the output signals Y1, Y2, and Y3 output from the learned model, the measured values ​​y1, y2, and y3 included in the evaluation data, and Figure 4BThe deviation is calculated using the formula shown on the lower side of , and the calculated deviation is stored in the storage unit 30 as the deviation 306. Since the deviation is calculated separately for the output signal of each learned model, the deviation 306 stored in the storage unit 30 includes the calculation results, that is, the deviations of a number equal to the number of learned models. The deviation calculated from the output signals Y1, Y2, Y3 output from each learned model and the measurement values ​​y1, y2, y3 obtained from the operating mechanical device 1 is used as an indicator for determining how much the state of the mechanical device 1 has changed from the initial stage of the normal state (that is, the normal period).

[0095] Next, in step S106 , the determination section 206 determines whether all the degrees of deviation calculated in step S105 are equal to or greater than a predetermined determination threshold value.

[0096] For example, the learned model 1 is a learned model that has learned the characteristic values ​​immediately after the start of use of the mechanical device 1 (i.e., at the first use), and thus the deviation calculated by using its output signal is used as an indicator for determining how much the state of the mechanical device 1 has changed from the start of use. In addition, the learned model 2 is a learned model that has learned the characteristic values ​​immediately after the first maintenance operation as will be described later, and thus the deviation calculated by using its output signal is used as an indicator for determining how much the state of the mechanical device 1 has changed from immediately after the first maintenance operation.

[0097] like Fig.11 As shown in , when all the deviations calculated by using all the stored post-learning models are equal to or greater than a predetermined determination threshold, it is determined that there is a sign of a fault. On the contrary, when at least one of the calculated deviations is less than the determination threshold, it is determined that there is no sign of a fault. It should be noted that although a different determination threshold may be set for each calculated deviation, as shown in FIG. Fig.11 As shown in , the determination can be made using the same determination threshold by normalizing the calculated deviation in advance. It should be noted that Fig.11 The following logic is shown: in this logic, the case where the normalized deviation is equal to or greater than the determination threshold is set to "true" represented by 1, the case where the normalized deviation is less than the determination threshold is set to "false" represented by 0, and when the logical product is "true" represented by 1, it is determined that there is a sign of a fault. Of course, a similar result can also be obtained from the following logic: in this logic, the case where the normalized deviation is equal to or greater than the determination threshold is set to "false" represented by 0, the case where the normalized deviation is less than the determination threshold is set to "true" represented by 1, and when the logical sum is "false" represented by 0, it is determined that there is a sign of a fault.

[0098] In the event that it has been determined in step S106 that there is no sign of a failure (ie, in the event of a NO result in step S106 ), the process returns to step S105 , and monitoring for a sign of a failure is continued.

[0099] On the contrary, in the case where it has been determined in step S106 that there is a sign of a failure (i.e., in the case where the result of step S106 is "yes"), the processing is transferred to step S107, and the notification section 207 transmits an alarm notification to the external device. As a method of notification, an appropriate mode such as lighting a lamp of the mechanical device, generating a warning sound, displaying a notification message on the display device 4 such as a display, or sending an e-mail can be adopted.

[0100] The user who has received the alarm notification indicating the sign of the failure in step S107 performs maintenance operation in step S108. Information related to the maintenance operation can be collected from the mechanical equipment 1 by the failure prediction device 2 as operation information 11, or input to the failure prediction device 2 by the user via the input device 3.

[0101] After the maintenance operation is performed in step S108, in step S109, the controller 20 determines whether to additionally generate a new learned model based on the data measured by the sensor group 10 after this maintenance operation and stores the new learned model in the learned model storage area 305. An event (such as a maintenance operation) that may cause the necessity of additionally generating a new learned model may also be referred to as a triggering event.

[0102] In the case where a new learned model is additionally generated (i.e., in the case where the result of step S109 is "yes"), the processing returns to step S102, the predetermined period after this maintenance operation is newly set as the normal period, and the processing of steps S102 to S104 is performed. That is, a new learned model is additionally generated and stored in the learned model storage area 305. Specifically, the extraction unit 203 extracts the data of the normal period after this maintenance operation from the feature value data 302, and stores the extracted data as additional learning data 307. The learned model adding unit 208 generates a learned model based on the additional learning data 307, and stores the generated learned model in the learned model storage area 305. Then, the processing is transferred to step S105, and the signs of failure are monitored by using all learned models including the added learned model.

[0103] In the case where a new post-learning model is not additionally generated (ie, in the case where the result of step S109 is “NO”), the process returns to step S105 via “NO” of step S101 , and the sign of a failure is monitored.

[0104] The above-mentioned processing flow will be described in more detail.

[0105] Figure 5 is a graph showing the degree of deviation with respect to time. By using the learned model 1 generated by learning the data of the "normal period" used for the first time, the degree of deviation is obtained from the output and the measured value of the learned model 1. Figure 5 In the example of , at time d1 after the start of use of the mechanical equipment or the like, the degree of deviation exceeds the determination threshold value of the failure prediction, and therefore it is determined that there is a sign of failure at this point in time.

[0106] When the maintenance operator of the mechanical equipment receives a notification of the presence of a sign of a failure, the maintenance operator performs maintenance operations on the mechanical equipment at a timing when the maintenance operations can be performed. Figure 5 The maintenance is performed at the time point described in the description. Although the condition of the mechanical equipment is improved by the maintenance operation, the measured value of the characteristic value is changed from the measured value of the characteristic value during the normal period of first use due to the replacement of parts, etc., so the deviation calculated after the maintenance operation does not necessarily return to the same level as the level during the normal period of first use. Figure 5 In the example of , at time point d2, the value of the deviation decreases compared to before maintenance, but is still greater than the determination threshold, so it is erroneously determined that there is a sign of failure, although the mechanical equipment is in good condition immediately after the maintenance operation.

[0107] Therefore, in this exemplary embodiment, a new learned model 2 is generated through machine learning using characteristic values ​​measured in the initial stage of the normal state after the maintenance operation as learning data, and the already generated learned model 1 and learned model 2 are used in combination for fault prediction.

[0108] Will refer to Figure 6 A mode is described in which a learned model 2 is newly generated and added after a maintenance operation, and a sign of a failure is detected by using the learned model 1 and the learned model 2 in combination.

[0109] Figure 6 The upper part of shows a graph showing the time change of the deviation obtained from the output and the measured value of each learned model. That is, the deviation M1 calculated by using the output of the learned model 1 and the deviation M2 calculated by using the output of the learned model 2 are plotted with respect to time. However, in the case of calculating the deviation by using the output of different learned models as in the present exemplary embodiment, the calculated deviation varies depending on the learning data used to learn the respective models, and therefore the deviations based on different learned models cannot be simply compared. Therefore, in Figure 6 In , the deviation is normalized so that its maximum value is equal when the learning data is given to the learned model as evaluation data.

[0110] Figure 6 The lower part shows a diagram in which the distribution of data is conceptually visualized by compressing the multidimensional data of eigenvalues ​​1 to 3 into two-dimensional data in dimension for the normal period t1 of the initial stage after the start of the operation, the period t2 immediately before the failure, and the normal period t3 of the initial stage after the maintenance operation. As a method for compressing into two-dimensional data, principal component analysis, etc., can be used, and the vertical axis and horizontal axis of the two-dimensional data represent the first principal component and the second principal component, respectively. Multiple pieces of data close in two-dimensional space can be determined as having a tendency to be close on the data of eigenvalues ​​1 to 3. It should be noted that the determination threshold value of the deviation degree used for calculation is shown as a circle in the figure of the two-dimensional data, and the size of the deviation degree is expressed by the distance from the center of the circle of the determination threshold value. However, these are not strict expressions and are conceptually used for convenience.

[0111] The data distribution of normal period t1 is Figure 6 The data of the normal period t3 is distributed in the area indicated by the "determination threshold for the degree of deviation calculated by using the output of the learned model 1". Figure 6 In the area indicated by "determination threshold for the degree of deviation calculated by using the output of the learned model 2". As described above, after the maintenance operation, although the device is in a normal state, the operating state may change due to replacement of parts, etc. and thus the range of data distribution may change.

[0112] The learned model 1 generated by machine learning of feature values ​​extracted from the data of the sensor group in the normal period t1 in the initial stage of the operation is a learned model that determines that the state within the data distribution range of the normal period t1 is a normal state in which there is no sign of a failure. In the period t2 immediately before the failure of the device and in which a notification of the presence of a sign of a failure should be issued, the deviation calculated by using the output of the learned model 1 is distributed in an area exceeding a certain threshold value, and therefore it is determined that there is a sign of a failure.

[0113] Although users Figure 6 A maintenance operation (such as replacement of a component) is performed at the time point of "maintenance" described in , but the state of the mechanical device does not return to the same state as the state at t1 when it was first used as described above. Therefore, although the device is in a good state immediately after the maintenance operation, the deviation M1 calculated by the output of the post-learning model 1 obtained by using the characteristic value measured by the input exceeds the determination threshold.

[0114] However, the learned model 2 generated by machine learning of the feature value of t3 which is the normal period immediately after the maintenance operation is a learned model that determines that the state within the data distribution range of the normal period t3 is a normal state in which there is no sign of a fault. Therefore, the deviation M2 calculated by using the output of the learned model 2 obtained by inputting the feature value measured immediately after the maintenance operation does not exceed the determination threshold.

[0115] Therefore, in the present exemplary embodiment, it is determined that there is no sign of a fault until all calculated deviations become equal to or greater than a determination threshold. In this case, it is determined that there is no sign of a fault until both the deviation M1 calculated by using the output of the learned model 1 and the deviation M2 calculated by using the output of the learned model 2 become equal to or greater than the determination threshold.

[0116] In the case where the learned models are generated separately by using the feature values ​​of the normal period t1 and the normal period t3 as in the present exemplary embodiment, Figure 6 As shown, the determination value can be easily set near the corresponding distribution area of ​​the characteristic value, so the determination accuracy of the symptom of the fault can be increased.

[0117] The case where a new post-learning model is additionally generated after a maintenance operation has been described here, and information related to the maintenance operation can be collected by the fault prediction device 2 from the mechanical equipment 1 as the operation information 11, or can be input to the fault prediction device 2 by the user via the input device 3.

[0118] It should be noted that, if Figure 7 As can be seen from the fact that step S109 is provided in the flowchart of , it is not necessary to add a new learned model every time a maintenance operation is performed. For example, in the case where the number of parts to be replaced in the maintenance operation is limited, there is a possibility that a determination of a symptom of a fault can be made with high accuracy after additionally generating a limited number of learned models without the need to additionally generate learned models.

[0119] For example, Fig. 10B An example of replacing a part in a maintenance operation of an injection molding machine is shown. In the maintenance operation of the injection molding machine, in order to minimize the time of operation stoppage, the part is replaced with a spare part, and thereafter the disassembled part is cleaned. Fig. 10BIn the example of , three screws are prepared and they are replaced in turn in each maintenance operation. It is assumed that the three screws have different characteristics within the range of a predetermined standard and all return to their original state after the cleaning operation. Although the learned model is added each time the screws are replaced in the maintenance operation, after the three learned models corresponding to the three screws are generated, it becomes unnecessary to additionally generate the learned model in the rotation of the screws in the maintenance operation. That is, the rotation of the screws in the subsequent maintenance operation is not a triggering event that causes the necessity to additionally generate a new learned model.

[0120] Next, a preferred method of providing information to a user in the failure prediction device serving as a control device and the failure prediction method serving as a control method of the present exemplary embodiment will be described.

[0121] Figure 8 , 9A 9B show an example of an image displayed on the screen of the display device 4 by the display section 210. By appropriately visualizing a part of the process of failure prediction, the user can intuitively grasp the control state.

[0122] Figure 8 The example of the display screen that is suitably used when, for example, the user desires to recognize the degree of deviation to grasp the state of the mechanical device 1 is shown. When the user specifies the time period for which the user wants to display the degree of deviation via the input device 3, the display unit 210 outputs the display screen information to the display device 4 to display the state of the mechanical device 1. Figure 8 In the display image, a display period field indicating a period designated by a user and a graph having a horizontal axis indicating date and time and a vertical axis indicating a degree of deviation are arranged.

[0123] By displaying such a graph, the user can easily grasp how much the operating state of the mechanical device 1 deviates from the normal state and its time change. In addition, if the value of the deviation at which the failure of the mechanical device occurs is roughly known based on past situations, the period until the failure can be estimated by looking at this graph. In addition, the cause of the failure can be inferred based on the speed of change of the state. For example, it is known that the deviation increases rapidly in the case of an event such as a mechanical rupture in a component, while the deviation increases gradually in the case of an event such as wear or dirtiness of a component in which the change occurs over time. By evaluating how the deviation changes, the cause of the failure that will occur next can be narrowed down to some extent.

[0124] Fig. 9A shows a screen for managing settings for learning models and is Figure 7An example of a display screen that is appropriately used in steps S102 to S104 (i.e., in each step for generating a learned model). In this screen, a field for inputting the name of a model as information for distinguishing the learned model to be generated from other learned models may be displayed. In addition, a feature value selection field for selecting feature values ​​to be used for the learning model may be displayed, and the feature values ​​to be used may be arbitrarily selected by the user by selecting a button displayed as "Use" or "Do not use" using a pointing device or the like. In addition, by entering a specific date and time in the normal period field, the range of learning data to be used for machine learning can be set.

[0125] By adopting such a screen configuration, even in the case where, for example, a feature value that can easily detect a sign differs for each mode in which a fault occurs, the user can easily generate a learned model with high detection accuracy for the sign of each fault mode.

[0126] Fig. 9B The figure shows a screen for managing the learned models displayed in the list, and is an example of a display screen including icons used appropriately for the user to newly add or delete learned models. Fig. 9B , the Add New Model button is displayed Fig. 9A The learning model setting screen is displayed, and the operation of adding the learned model can be started. In addition, in the case where, for example, a learned model is added by mistake, the user can select Fig. 9B In the present exemplary embodiment, even in the case where a learned model is additionally generated by machine learning due to an error, as long as the unnecessary learned model is deleted, it is unnecessary to learn again, so this screen is actually useful for the user.

[0127] As described above, in the present exemplary embodiment, in the case where the change pattern of the characteristic value to be evaluated deviates from the change pattern of the characteristic value predicted by all the learned models beyond a predetermined range, it is determined that the normal state of the mechanical device 1 will not continue for a long time. In other words, in the case where the degree of deviation of the output of all the learned models from the measured characteristic value is equal to or greater than a predetermined determination threshold, it is predicted that the occurrence of a failure of the mechanical device 1 is approaching, and it is determined that a maintenance operation should be performed.

[0128] By adding a learning model according to a change in the operating state of the device as in the present exemplary embodiment, fault prediction can be performed with high accuracy even in a case where it is difficult to collect data representing all normal states in advance. In addition, since the normal data learned by a machine-learned model is limited to an operating state of the device, the complexity of the data input in the machine-learned model does not change. Therefore, an event such as a decrease in detection accuracy caused by insufficient adjustment of the parameters of the machine-learned model is unlikely to occur. That is, it is possible to detect signs of a fault with high accuracy without excessively increasing the operating load on the system.

[0129] It is to be noted that the present invention is not limited to the above-described exemplary embodiments or examples, and may be modified in various ways within the technical concept of the present invention.

[0130] Mechanical equipment that can be the subject of failure prediction is not limited to the injection molding machine, and failure prediction can be performed for various mechanical equipment that performs replacement or cleaning of mechanical parts in maintenance operations, such as injection machines for powders and liquids.

[0131] Other embodiments

[0132] Embodiments of the present invention may also be implemented by reading out and executing computer executable instructions (e.g., one or more programs) recorded on a storage medium (which may also be more completely referred to as a 'non-transitory computer-readable storage medium') to perform one or more of the functions of the above-mentioned embodiments and / or a computer of a system or device including one or more circuits (e.g., an application-specific integrated circuit (ASIC)) for performing one or more of the functions of the above-mentioned embodiments, and by a computer of the system or device, for example, reading out and executing computer executable instructions from a storage medium to perform one or more of the functions of the above-mentioned embodiments and / or controlling one or more circuits to perform one or more of the functions of the above-mentioned embodiments. The computer may include one or more processors (e.g., a central processing unit (CPU), a microprocessing unit (MPU)), and may include a network of separate computers or separate processors to read out and execute computer executable instructions. Computer executable instructions may be provided to the computer, for example, from a network or a storage medium. The storage medium may include, for example, a hard disk, a random access memory (RAM), a read-only memory (ROM), a storage device of a distributed computing system, an optical disk (such as a compact disk (CD), a digital versatile disk (DVD), or a Blu-ray disk (BD)) TM ), one or more of flash memory devices, memory cards, etc.

[0133] The embodiments of the present invention may also be implemented by providing software (program) for performing the functions of the above-described embodiments to a system or device via a network or various storage media, and a computer or a central processing unit (CPU) or a microprocessing unit (MPU) of the system or device reads and executes the program.

[0134] While the present invention has been described with reference to exemplary embodiments, it is to be understood that the invention is not limited to the disclosed exemplary embodiments.The scope of the following claims is to be accorded the broadest interpretation so as to encompass all such modifications and equivalent structures and functions.

Claims

1. An information processing method for a controller to determine a state of a mechanical device capable of mounting a first component or a second component, the information processing method comprising: a measuring step, in which the controller obtains a measurement value of a sensor provided in the mechanical device; a first storage step in which the controller generates a first model by machine learning using a measurement value of the sensor measured during a first period in which the mechanical device keeps the first component mounted thereon, and stores the first model in a storage unit; a second storage step in which the controller generates a second model different from the first model through machine learning using a measurement value of the sensor measured in a second period after the first component mounted on the mechanical equipment is replaced with the second component, and stores the second model in the storage unit; a determination step in which the controller determines the state of the mechanical device by using the measurement values ​​of the sensor measured in an evaluation period and the first model and the second model stored in the storage section, and wherein the controller controls the state of the storage unit so as to maintain the first model corresponding to the first component generated in the first storage step, and The controller also controls a state of the storage section so as to add and maintain the second model corresponding to the second component generated in the second period.

2. The information processing method according to claim 1, further comprising an extraction step in which the controller extracts a feature value from the measurement value of the sensor measured in time series, in, In the first storage step, the controller generates the first model by machine learning using feature values ​​extracted from measurement values ​​measured in the first period, and stores the first model in the storage section, and Among them, in the second storage step, the controller generates the second model through machine learning using feature values ​​extracted from measurement values ​​measured in the second period, and stores the second model in the storage unit.

3. The information processing method according to claim 2, wherein: In the determination step, the controller inputs feature values ​​extracted from the measurement values ​​measured in the evaluation period into all the models stored in the storage section, and determines the state of the mechanical device by using output signals output from the respective models.

4. The information processing method according to claim 3, wherein In the determining step, The controller calculates a deviation of the output signal from a measurement value measured in the evaluation period, and The controller determines whether all calculated degrees of deviation are equal to or greater than a predetermined determination threshold.

5. The information processing method according to claim 1, wherein the mechanical device is an injection molding machine including screws, wherein the sensor is a sensor configured to measure at least one of a current, a position, a vibration, and a sound of an actuator driving the screw, and The controller inputting characteristic values ​​of the measurement values ​​of the sensor corresponding to a predetermined number of timings consecutive in time series into a predetermined function, obtaining an estimated feature value corresponding to a timing subsequent to the timing, the estimated feature value being estimated by using the predetermined function, obtaining a measured characteristic value of a measurement value actually measured by the sensor at the subsequent timing, and The first model or the second model is generated by repeatedly updating the predetermined function so that the difference between the estimated eigenvalue and the measured eigenvalue becomes smaller. 6 . The information processing method according to claim 1 , wherein a predetermined period of time after the mechanical device starts to be used is set as the first period of time.

7. The information processing method according to any one of claims 1 to 5, wherein a predetermined period after use of the mechanical device starts after the first component mounted on the mechanical device is replaced with the second component is set as the second period.

8. The information processing method according to claim 4, wherein: In the determining step, the controller determines that there is a sign of a failure of the mechanical device in a case where all the calculated deviations are equal to or greater than the predetermined determination threshold value.

9. The information processing method according to any one of claims 1 to 5, further comprising a notification step in which a determination result of the determination step is notified to a user.

10. The information processing method according to claim 1, further comprising an information display step, in which the controller displays information of at least one of the measuring step, the first storage step, the second storage step, and the determining step on a display device.

11. The information processing method according to claim 10 further includes a model information display step, in which, when the first component installed on the mechanical equipment is replaced by the second component, the controller displays information of the model stored in the storage unit and / or an icon for indicating a newly added model on the display device.

12. The information processing method according to claim 11, wherein: In the model information display step, the controller displays information related to selection of a feature value, information related to designation of the first period or the second period, and / or information distinguishing the second model from other models on the display device.

13. The information processing method according to claim 4, further comprising a deviation display step, in which the controller displays the deviation on a display device in a time series.

14. An information processing device comprising a controller, The controller is configured as obtaining a measurement value of a sensor provided in a mechanical device capable of mounting the first component or the second component; generating a first model by machine learning using the measurement value of the sensor measured during a first period in which the mechanical device keeps the first component mounted thereon, and storing the first model in a storage unit; generating a second model different from the first model through machine learning using a measurement value of the sensor measured in a second period after the first component mounted on the mechanical equipment is replaced with the second component, and storing the second model in the storage unit; determining a state of the mechanical device by using the measurement values ​​of the sensor measured in an evaluation period and the first model and the second model stored in the storage section; controlling a state of the storage unit so as to maintain the generated first model corresponding to the first component; as well as The state of the storage section is controlled so as to add and hold the generated second model corresponding to the second component. 15 . The information processing apparatus according to claim 14 , wherein the controller notifies a user of a result of determination of the state of the mechanical device.

16. The information processing device according to claim 14, further comprising a display device, in, When the controller performs at least one of the following processes, the controller displays information of the at least one process on the display device: a process of obtaining the measurement value of the sensor, a process of generating the first model and storing the first model in the storage unit, a process of generating the second model and storing the second model in the storage unit, and a process of determining the state of the mechanical equipment.

17. The information processing device according to claim 16, wherein: In a case where the first component mounted on the mechanical equipment is replaced with the second component, the controller displays information of the model stored in the storage unit and / or an icon indicating a newly added model on the display device.

18. The information processing device according to claim 17, wherein the controller displays on the display device information related to selection of feature values, information related to designation of the first and second time periods, and / or information distinguishing the second model from other models.

19. The information processing device according to claim 14, further comprising a display device, wherein the controller extracts a feature value from the measurement value of the sensor measured in the evaluation period, inputs the feature value into the first model and the second model stored in the storage unit, and calculates the degree of deviation of the output signal from the measurement value measured in the evaluation period by using the output signal output from the first model and the second model, and The controller displays the degree of deviation on the display device in time series. 20 . A method of manufacturing a product, wherein the product is manufactured by using the information processing apparatus and a mechanical device according to claim 14 . 21 . A non-transitory computer-readable recording medium storing a control program that enables execution of the information processing method according to claim 1 .

Images