JP2019049940A - Abnormal process estimating method, abnormal process estimating apparatus, and abnormal process estimating program - Google Patents

Abstract

To estimate, with satisfactory accuracy, an abnormal process which can be a factor of abnormality in a quality of a product.SOLUTION: An abnormal process estimating method according to the present embodiment has a collecting step, a generating step, an evaluating step, and a determining step. In the collecting step, a plurality of sensors are provided to collect sensor data of the sensors in a manufacturing line consisting of a plurality of processes. In the generating process, based on sensor data regarding normal quality in a normal state of the manufacturing line, machine learning is carried out so that lapsed time in the above process should be included in a feature vector, thereby generating a plurality of correlation models among the sensors corresponding to the respective processes. In the evaluating step, an evaluation amount of each process obtained in consideration of sensor data at an arbitral evaluation and lapsed time of the process with respect to the sensor data is input to each of corresponding correlation models, and based on an output value of the correlation model thus obtained, deviation from the normal state with respect to each of the processes is evaluated. In the determining step, an abnormal process which can be a factor of abnormality in a product quality is determined from the processes based on the above deviation.SELECTED DRAWING: Figure 2

Description

  Embodiments disclosed herein relate to an abnormal process estimation method, an abnormal process estimation device, and an abnormal process estimation program.

  2. Description of the Related Art Conventionally, there is known a technique for predicting the occurrence of a failure by detecting a failure sign by monitoring sensor values of sensors provided in the machinery and equipment (see, for example, Patent Document 1).

  The technology disclosed in Patent Document 1 compares the sensor value of a sensor mounted on a vehicle with a normal threshold value, and determines whether or not an abnormality occurs in the sensor based on the comparison result. Then, when it is determined that an abnormality occurs in the sensor, an evaluation index for evaluating a failure sign of the sensor using the integrated value of the difference between the sensor value and the normal threshold, the abnormality occurrence duration time, etc. To calculate the failure index of the sensor using such evaluation index.

  Such a technology is considered to be applicable not only to a vehicle, but also to, for example, a manufacturing apparatus that performs a predetermined process to manufacture a product.

JP 2011-230634 A

  However, the above-mentioned prior art has room for further improvement in accurately estimating an abnormal process which causes a quality abnormality of a product.

  Specifically, although the above-mentioned prior art is suitable for detecting a clear abnormal condition exceeding the threshold value, the behavior of the entire system may appear unclear due to, for example, aging of each component. The sensor value up to the threshold is not suitable for detecting an unclear abnormal state which does not indicate the sensor value.

  Therefore, when the above-described conventional technology is applied to a manufacturing apparatus or the like, even if the product has a quality abnormality due to an unclear abnormality, the factor can not be detected. This is because, if the production method of the product is a line production method etc., the product is manufactured through a series of plural steps by plural manufacturing devices, so if the above-mentioned quality abnormality occurs, from among the plural steps It means that the abnormal process which became the factor can not be estimated accurately.

  One aspect of the embodiment is made in view of the above, and is an abnormal process estimation method, an abnormal process estimation device, and an abnormal process estimation program capable of accurately estimating an abnormal process that is a factor of product quality abnormality. Intended to provide.

  The abnormal process estimation method according to an aspect of the embodiment includes a collection process, a generation process, an evaluation process, and a determination process. The collecting step includes a plurality of sensors, and collects sensor data of the sensors in a manufacturing line including a plurality of steps. The generation step performs machine learning by performing machine learning such that an elapsed time between the steps is included in a feature vector based on the sensor data of normal quality while the production line is in a normal state. Generate a plurality of correlation models between the sensors corresponding to The evaluation step is obtained by inputting the sensor data at any time of evaluation and an evaluation component for each step in which an elapsed time between the steps is added to the sensor data to the corresponding correlation model. The degree of deviation from the normal state in each of the steps is evaluated based on the output value of the correlation model. The determination process determines an abnormal process that causes a quality abnormality of the product among the processes based on the degree of deviation.

  According to one aspect of the embodiment, it is possible to accurately estimate an abnormal process that causes a quality abnormality of a product.

FIG. 1A is a schematic explanatory view (part 1) of the abnormal process estimation method according to the embodiment. FIG. 1B is a schematic explanatory view (No. 2) of the abnormal process estimation method according to the embodiment. FIG. 1C is a schematic explanatory view of an abnormal process estimation method according to a comparative example. FIG. 1D is a schematic explanatory view (No. 3) of the abnormal process estimation method according to the embodiment. FIG. 2 is a block diagram of the abnormal process estimation system according to the embodiment. FIG. 3A is a process explanatory diagram of preprocessing (part 1). FIG. 3B is a process explanatory diagram of the preprocessing (part 2). FIG. 3C is a process explanatory diagram of the preprocessing (part 3). FIG. 3D is a processing explanatory view of the preprocessing (part 4). FIG. 4A is a process explanatory diagram of the generation process. FIG. 4B is a process explanatory diagram of the evaluation process. FIG. 5A is a diagram showing a verification result according to the present embodiment. FIG. 5B is a diagram showing a verification result according to the comparative example. FIG. 6 is a flowchart showing a processing procedure executed by the abnormal process estimation device. FIG. 7 is a hardware configuration diagram showing an example of a computer that implements the function of the abnormal process estimation device.

  Hereinafter, embodiments of an abnormal process estimation method, an abnormal process estimation device, and an abnormal process estimation program disclosed in the present application will be described in detail with reference to the attached drawings. Note that the present invention is not limited by the embodiments described below.

  Moreover, below, the case where the manufacturing line by which the several process was serialized with the several manufacturing apparatus becomes object of abnormal process estimation is demonstrated. Also, when the manufacturing apparatus is generically referred to, "manufacturing apparatus 100" is described, and when a different manufacturing apparatus 100 is described, "-n" (n: Numbered in the form of natural numbers).

  First, an outline of the abnormal process estimation method according to the present embodiment will be described with reference to FIGS. 1A to 1D. FIG. 1A, FIG. 1B and FIG. 1D are outline | summary explanatory drawing (the 1)-(the 3) of the abnormal process estimation method which concerns on embodiment. FIG. 1C is a schematic explanatory view of an abnormal process estimation method according to a comparative example.

  In the description using FIGS. 1A and 1B, the manufacturing apparatus 100 alone is taken as an example, and the basic concept of the abnormal process estimation method according to the present embodiment when one step is performed by the manufacturing apparatus 100. explain. As shown to FIG. 1A, the manufacturing apparatus 100 is provided with the sensor group of sensor S-1-S-n.

  And the abnormal process estimation method according to the present embodiment is based on the sensor data from the sensor group, the correlation between the sensors S-1 to S-n (hereinafter sometimes referred to as "sensor correlation") It grasps, and grasps the change of the behavior of the manufacturing apparatus 100 whole based on the change of the sensor correlation.

  Specifically, as shown in FIG. 1B, in the abnormal process estimation method according to the present embodiment, the machine data is used for each sensor S-1 to S-n showing sensor correlativity of "normal quality". Learning is performed (step S1) to generate a correlation model 12d '. In the present embodiment, deep learning is used as a machine learning algorithm. Since deep learning is a known method, detailed description here is omitted.

  The “normal quality portion” refers to a sensor data group when the manufacturing apparatus 100 manufactures a product with no quality abnormality, for example, a sensor data group accumulated for a predetermined number of products manufactured from the first operation. . By means of the correlation model 12 d ′ generated based on the sensor correlation of the normal quality, it is possible to model the normal state of the manufacturing apparatus 100, that is, manufacturing a product without quality abnormality.

  Then, in the abnormal process estimation method according to the present embodiment, as shown in FIG. 1B, the sensor data of each of the sensors S-1 to S-n showing the sensor correlation of “estimated portion” with respect to the correlation model 12d ′. Is input, and the shift amount of the correlation is calculated from the output value (regression value) of the correlation model 12 d ′ obtained as a result. Then, the degree of deviation from the normal state is evaluated based on the amount of deviation of the correlation (step S2).

  If the deviation degree is large enough to exceed the predetermined threshold, the manufacturing apparatus 100 indicates a failure sign, and an abnormal process can be estimated as a product abnormality may occur due to the process performed by the manufacturing apparatus 100. Here, the "evaluation portion" refers to, for example, a sensor data group output in real time from the manufacturing apparatus 100, or a sensor data group at the time of manufacturing the corresponding product when a quality abnormality is recognized in the product in a test or the like. .

  As described above, in the abnormal process estimation method according to the present embodiment, the behavior of the entire manufacturing apparatus 100 is grasped by sensor correlativity, and the change in the behavior is the correlation model 12d ′ that models the sensor correlativity in the normal state. It can be obtained by the output value. Then, the magnitude of deviation from the normal state based on the output value indicates that the manufacturing apparatus 100 has a sign of failure, and it is presumed that the process performed by the manufacturing apparatus 100 is an abnormal process that causes quality abnormality. Do.

  Therefore, according to the abnormal process estimation method according to the present embodiment, all abnormal process estimation methods that are required, for example, can be used as long as an abnormal process estimation method using a failure model generated by machine learning based on sensor data at the time of failure occurrence. There is no need for complicated processes such as the modeling of failure phenomena. Therefore, while being able to catch the failure sign of the manufacturing apparatus 100 simply and accurately, it becomes possible to estimate the abnormal process which causes the quality abnormality accurately.

  In the abnormal process estimation method according to the present embodiment, as shown in FIG. 1A, the temporal variation of sensor data of each of the sensors S-1 to S-n, that is, the time-series correlation is a feature vector (hereinafter simply referred to as “ The machine learning included in “vector”) is performed to make it possible to grasp the time-series correlation of sensor data. This can contribute to accurately grasping the failure sign of the manufacturing apparatus 100.

  Here, on the premise of the above-described basic idea, the case of the manufacturing line 2 in which a plurality of processes are serialized by a plurality of manufacturing apparatuses 100 will be considered. As shown in FIG. 1C, the manufacturing line 2 includes manufacturing devices 100-1 and 100-2.

  The manufacturing apparatuses 100-1 and 100-2 both have a sensor group, and for example, the manufacturing apparatus 100-1 executes a process of 2t in the order of the first process and the second process, and the manufacturing apparatus 100-2 performs the second process. It is assumed that the third step is performed based on the result of the step.

  When applying the basic idea described above to the manufacturing line 2, the sensor data of each process of the first to third processes are input as a learning data set d for learning, and this is used as a batch machine learning It is conceivable to execute (collective learning) and generate a correlation model 12 d ′.

  However, when multiple steps are performed as shown in FIG. 1C, an elapsed time is usually present between each step. For example, when an adhesive, a cleaning agent, or the like is used in a certain process, the elapsed time is a factor that changes the degree of dryness or the like. In addition, for example, when heating, heat retention, cooling, or the like is performed in a certain process, the elapsed time is also a factor such as temperature fluctuation. That is, when the entire production line 2 is regarded as one manufacturing device 100, the elapsed time can be one of vectors indicating the feature of the time-series correlation in the manufacturing device 100.

  In addition, in each process, the contents of the process to be performed are different. For example, when a plurality of processes are performed by one manufacturing apparatus 100, there is a process in which a process highly relevant to the sensor group is performed. It is conceivable that there is a process in which processing that is less related to the sensor group is performed. If the normal state of the entire manufacturing line 2 in which each of these steps is performed is modeled by one correlation model 12d ', the above-mentioned correlation shift amount may be derived more than necessary.

  Therefore, as shown in FIG. 1D, in the abnormal process estimation method according to the present embodiment, while basically setting sensor data for each process as a learning target, a learning data set d is prepared in which the elapsed time between processes is taken into consideration. The correlation model group 12d is generated by executing machine learning (division learning) divided for each learning data set d.

  As shown in FIG. 1D, for example, in split learning in which the first step is a learning target, a first correlation model 12d-1 is generated. In addition, in divided learning in which the elapsed time between the first and second steps and the second step are learning targets, a second correlation model 12d-2 is generated. In addition, in divided learning in which the elapsed time between the second and third steps and the third step are learning targets, a third correlation model 12d-3 is generated.

  Then, in the evaluation stage, after preparing a data set taking into account the elapsed time also for the data set for evaluation, the evaluation data set for the first step is the first correlation model 12d-1, and it is between the first and second steps. The evaluation data set of the second step and the elapsed time of the second step are evaluated by the second correlation model 12d-2, and the evaluation data set of the second and third steps and the evaluation step of the third step are evaluated by the third correlation model 12d-3. As mentioned, it was decided to evaluate the degree of deviation from the normal state while switching the model.

  This makes it possible to calculate the degree of deviation from the normal state with high accuracy for each process to be evaluated. Then, based on the calculated degree of deviation from the normal state, in the abnormal process estimation method according to the present embodiment, the position of the abnormal process is estimated based on the time position with the large degree of deviation, for example, based on sensor data corresponding to the position Calculate the contribution order of each sensor. The contribution order corresponds to the location of each sensor, etc., which can help to identify the location where there is a failure sign.

  As described above, in the abnormal process estimation method according to the present embodiment, machine learning is performed based on the learning data set d which is divided for each process and in which the elapsed time between processes is added. A correlation model group 12d including correlation models 12d-1, 12d-2, 12d-3 is generated.

  Then, in the abnormal process estimation method according to the present embodiment, the evaluation data set which is similarly divided for each process and in which the elapsed time between processes is added is switched between the correlation model group 12d while changing for each process to be evaluated. Input to the corresponding correlation model, and evaluate the degree of deviation from the normal state for each process based on the obtained output value.

  Then, in the abnormal process estimation method according to the present embodiment, as the evaluation result, the abnormal process is estimated from the time position at which the degree of deviation from the normal state is equal to or greater than a predetermined threshold. Therefore, according to the abnormal process estimation method according to the present embodiment, it is possible to accurately estimate an abnormal process that causes a quality abnormality of a product.

  Hereinafter, the configuration of the abnormal process estimation system 1 to which the above-described abnormal process estimation method is applied will be described more specifically.

  FIG. 2 is a block diagram of the abnormal process estimation system 1 according to the present embodiment. In addition, in FIG. 2, the component required in order to demonstrate the characteristic of this embodiment is represented with a functional block, and the description about a general component is abbreviate | omitted.

  In other words, each component illustrated in FIG. 2 is functionally conceptual and does not necessarily have to be physically configured as illustrated. For example, the specific form of distribution and integration of each functional block is not limited to that shown in the drawings, and all or part of them may be functionally or physically distributed in any unit depending on various loads, usage conditions, etc. It is possible to integrate and configure.

  In the description using FIG. 2, the description may be simplified or omitted for the components already described.

  As shown in FIG. 2, the abnormal process estimation system 1 includes an abnormal process estimation device 10 and a manufacturing line 2. The manufacturing line 2 includes a plurality of manufacturing apparatuses 100. Each of the plurality of manufacturing apparatuses 100 includes a sensor group.

  The abnormal process estimation device 10 and the manufacturing device 100 are connected in a network and communicably provided, and the abnormal process estimation device 10 is provided so as to be able to appropriately collect sensor data from the manufacturing device 100.

  The abnormal process estimation device 10 includes a control unit 11 and a storage unit 12. The control unit 11 includes a collection unit 11a, a preprocessing unit 11b, an extraction unit 11c, a generation unit 11d, an evaluation unit 11e, a determination unit 11f, and a notification unit 11g.

  The storage unit 12 is a storage device such as a hard disk drive, a non-volatile memory, and a register, and includes a collection log 12a, a collective log group 12b, a divided learning data set 12c, a correlation model group 12d, and an evaluation data set 12e. The evaluation information 12f is stored. The evaluation information 12 f includes a divergence degree 12 fa and a contribution order 12 fb.

  The control unit 11 performs overall control of the abnormal process estimation device 10. The collection unit 11a collects sensor data from the sensor group of the manufacturing apparatus 100 at a predetermined cycle, and stores the data in the collection log 12a.

  The preprocessing unit 11b generates a collective log representing a series of processes in the manufacturing line 2 based on the collection log 12a as preparation for machine learning and evaluation to be executed later, and accumulates the collective log in the collective log group 12b. The collective log includes contents corresponding to the elapsed time between processes, and one collective log corresponds to one series of processes performed on the production line 2.

  Here, the preprocessing performed by the preprocessing unit 11b will be specifically described with reference to FIGS. 3A to 3D. FIGS. 3A to 3D are process explanatory views (part 1) to (part 4) of preprocessing.

  As shown to FIG. 3A, as an example, the manufacturing line 2 shall consist of 5 processes, a 1st-3rd process shall be a manufacturing process, and a 4th and 5th process shall be a test process. Note that this example does not limit the total number of processes in the manufacturing line 2, the number of manufacturing processes, and the number of test processes.

  In the case of the production line 2, the collection log 12a includes, as the "production log", the "first log" for the first process, the "second log" for the second process, and the "third log" for the third process. Is stored as one unit of a series of process steps. Further, as the "test log", the "fourth log" for the fourth process and the "fifth log" for the fifth process are accumulated as one unit of a series of processes.

  The first to fifth logs each include a "standard part" and a "dedicated part". Here, as shown in FIG. 3B, the “standard part” stores, for example, a unique number for each product to be manufactured, a date, time, alarm information, and the like.

  Further, as shown in FIG. 3C, for example, each sensor data being manufactured is stored in the “dedicated part” of the manufacturing log. Each sensor data is, for example, position information, speed information, torque information, and the like. Further, as shown in FIG. 3D, for example, test results and the like are stored in the “dedicated part” of the test log.

  It returns to the explanation of FIG. 3A. The preprocessing unit 11 b generates a collective log of the entire manufacturing process in which one series of process steps is one unit based on the collection log 12 a (step S 3). At this time, the preprocessing unit 11 b takes into account the elapsed time between processes in the collective log.

  Specifically, as shown in FIG. 3A, the collective log includes “standard part”, “first log dedicated part”, “second log dedicated part”, and “third log dedicated part”. The contents of the first to third logs in the collection log 12a are stored in chronological order of execution of each process. For example, the contents of the "dedicated part" of the first log are stored in the "first log dedicated part" of the row corresponding to the first process (see the M1 part in the figure). "0" is stored in the "second log dedicated unit" and the "third log dedicated unit".

  Similarly, the contents of the "dedicated part" of the second log are stored in the "second log dedicated part" of the row corresponding to the second process (see the M2 part in the figure). “0” is stored in the “first log dedicated unit” and the “third log dedicated unit”. Similarly, the contents of the "dedicated part" of the third log are stored in the "third log dedicated part" of the row corresponding to the third process (see the M3 part in the figure). “0” is stored in the “first log dedicated unit” and the “second log dedicated unit”.

  Then, in the pre-processing unit 11b, all the dedicated portions are zero-filled on positions corresponding to the elapsed time between the first and second steps and the elapsed time between the second and third steps in time series (that is, The lines (filled with “0”) are inserted for the number of lines corresponding to the elapsed time between steps (step S31). That is, in the example shown in FIG. 3A, the elapsed time between the second and third steps is equivalent to twice the elapsed time between the first and second steps. The preprocessing unit 11 b accumulates the collective log generated in this manner in the collective log group 12 b.

  It returns to the explanation of FIG. The extraction unit 11c divides and extracts the learning data set d for normal quality corresponding to the normal operation of the manufacturing line 2 from the collective log group 12b for each process to be learned, and stores the learning data set d in the divided learning data set 12c. .

  Further, the extraction unit 11 c extracts a data set for evaluation corresponding to the time of evaluation from the collective log group 12 b and stores the data set in the evaluation data set 12 e.

  The generation unit 11 d performs machine learning by deep learning using the divided learning data set 12 c as an input, and generates a correlation model group 12 d.

  The evaluation unit 11e inputs the evaluation data set 12e extracted by the extraction unit 11c to the corresponding correlation models 12d-1, 12d-2, 12d-3 of the correlation model group 12d, and outputs the output results from the correlation models. receive.

  Then, the evaluation unit 11e calculates various evaluation values for determining a failure sign based on the received output result, and stores the evaluation values in the evaluation information 12f. The evaluation value corresponds to the deviation degree 12fa and the contribution order 12fb.

  Specifically, the evaluation unit 11e inputs the evaluation data set 12e to the corresponding correlation models 12d-1, 12d-2, 12d-3 of the correlation model group 12d, and the correlation models 12d-1, 12d-2 , 12d-3, the degree of deviation from the normal state corresponding to each of the correlation models 12d-1, 12d-2, 12d-3 is calculated. Further, the evaluation unit 11e stores the calculated degree of deviation in the degree of deviation 12fa of the evaluation information 12f.

  Further, the evaluation unit 11 e calculates the contribution order of each sensor based on the sensor data of each sensor, and stores the contribution order in the contribution order 12 fb of the evaluation information 12 f.

  Here, the generation process performed by the generation unit 11 d and the evaluation process performed by the evaluation unit 11 e will be described more specifically with reference to FIGS. 4A and 4B. FIG. 4A is a process explanatory diagram of the generation process. Moreover, FIG. 4B is process explanatory drawing of evaluation processing.

  As shown in FIG. 4A, in the collective log group 12b, collective logs corresponding to normal quality are accumulated. The batch log of the normal quality is extracted while being “divided” by the extraction unit 11 c described above, and is stored as the divided learning data set 12 c.

  In the case of the example corresponding to FIG. 3A already described, as shown in FIG. 4A, in the divided learning data set 12c, the learning data set d1 corresponding to “the first step” and “between the first and second steps” And a learning data set d3 corresponding to "the elapsed time between the second and third steps + the third step".

  Then, the generation unit 11d executes "machine learning" based on the learning data set d1 to generate a first correlation model 12d-1. Further, the generation unit 11 d executes “machine learning” based on the learning data set d2 to generate a second correlation model 12 d-2. Further, the generation unit 11 d executes “machine learning” based on the learning data set d3 to generate a third correlation model 12 d-3.

  Then, for each correlation model 12d-1, 12d-2, 12d-3 of the correlation model group 12d generated in this way, the evaluation unit 11e is extracted by the above-mentioned extraction unit 11c as shown in FIG. 4B. Also, "input" the evaluation data set 12e corresponding to the collective log for the evaluation.

  At this time, the evaluation unit 11 e inputs each of the correlation models 12 d-1, 12 d-2, and 12 d-3 to be input destinations while “switching for each process to be evaluated” and inputs each correlation model 12 d-1, 12 d- The degree of deviation from the normal state corresponding to each of the correlation models 12d-1, 12d-2, and 12d-3 is calculated based on the output values from 2, 12d-3.

  The determination unit 11 f described later determines that there is a failure sign at a time position where the degree of deviation from the normal state calculated by the evaluation unit 11 e is equal to or more than a predetermined threshold, and estimates that it corresponds to “quality abnormality”. Become.

  It returns to the explanation of FIG. The determination unit 11 f refers to the degree of deviation 12 fa, and when the degree of deviation from the normal state corresponding to any of the correlation models 12 d-1, 12 d-2, and 12 d-3 is equal to or greater than a predetermined threshold, It is determined that there is a notification request to the notification unit 11g.

  The notification unit 11g notifies an external device of alert notification when it receives a notification request instruction of failure indication from the determination unit 11f. Further, at this time, the notification unit 11g refers to the deviation degree 12fa and the contribution order 12fb, for example, the process position estimated from the deviation degree from the normal state and the corresponding time position, and each sensor S- of the contribution order superior. The names of 1 to S-n and the locations can be notified together.

  The user can know, for example, the degree of abnormality and the abnormal process position that may have caused the quality abnormality by confirming the deviation from the normal state and the process position corresponding thereto. Further, for example, by checking a sensor at the top of the contribution order, it is possible to estimate the occurrence location and the cause of the abnormality.

  Next, verification examples according to the verification example and the comparative example according to the present embodiment are shown in FIGS. 5A and 5B. FIG. 5A is a diagram showing a verification result according to the present embodiment. Moreover, FIG. 5B is a figure which shows the verification result concerning a comparative example.

  In FIG. 5A and FIG. 5B, the case where the manufacturing line 2 bonds the support substrate W2 to the to-be-processed substrate W1 is mentioned as an example. In recent years, the diameter and thickness of the target substrate W1 such as a silicon wafer and a compound semiconductor wafer have been increased, and the target substrate W1 may be warped or broken during transportation or polishing processing. Therefore, the substrate to be processed W1 is reinforced by bonding a support substrate W2 such as a glass substrate to make a product.

  The production line 2 includes a coating device, a bonding device, and an inspection device, as shown in FIG. 5A. The coating apparatus and the bonding apparatus are an example of the manufacturing apparatus 100 responsible for the manufacturing process, and the inspection apparatus is an example of the manufacturing apparatus 100 responsible for the test process.

  As shown in FIG. 5A, in the production line 2, "adhesion agent discharge", "adhesive application", "cleansing agent injection", and "peripheral part cleaning" are performed by the bonding device in order from component supply in the coating device. "Heating" and "crimping" are performed on the inspection apparatus.

  In the “adhesive discharge”, the adhesive is discharged onto the processing target substrate W1, and the “discharged state” is observed as an analysis target by an image or the like. In “adhesive application”, the adhesive is applied to the surface of the substrate W1 by centrifugal force by rotating the substrate W1 to which the adhesive is discharged. Sensor data "position information", "speed information", "torque information", etc. are collected as an analysis target.

  In the "cleaning agent injection", the cleaning agent is injected to the peripheral portion of the processing target substrate W1 which does not require the adhesive, and the "injection state" is observed as an analysis target by an image or the like. In the “peripheral part cleaning”, the peripheral part of the substrate to be processed W1 is cleaned by the cleaning agent by rotating the substrate to be processed W1. Sensor data "position information", "speed information", "torque information", etc. are collected as an analysis target.

  In "heating", the substrate to be processed W1 is heated in order to volatilize the organic solvent added to the adhesive and to increase the viscosity of the adhesive. As a subject of analysis, "heating time", "warming time" and the like are observed. In "crimping", the supporting substrate W2 is positioned at the same horizontal position different from the substrate to be processed W1 in height, and the two substrates W1 and W2 are bonded by crimping from the vertical direction. Sensor data "position information", "speed information", "torque information", etc. are collected as an analysis target. In the "test", the bonded products are tested.

  Then, in the verification example according to the present embodiment, for example, the “adhesive application” in the coating apparatus is set as the “first step”, and the time from the end of the “adhesive application” to the end of the “peripheral part cleaning” is “elapsed time + 2nd process ”. In addition, from the end of the "peripheral part cleaning" to the end of the "crimping" is referred to as "elapsed time + third step".

  Then, each of the correlation models 12d-1, 12d-2, 12d- and the learning data set d of the normal quality for each of the "first step", "elapsed time + second step" and "elapsed time + third step" The verification result when 3 is generated and the training data d of normal quality is input to these and autoregressed, and the verification result when the data set for abnormal quality when the quality abnormality is output are input Obtained.

  As shown in FIG. 5A, according to the verification result of “at normal quality part input time”, the degree of deviation from the normal state is generally low, and each correlation model 12d-1, 12d-2, 12d− It can be seen that 3 can correctly model production line 2 in a normal state.

  Moreover, according to the verification result of "at the time of the input of abnormal quality", there is a time position indicating the degree of deviation from the normal state at the predetermined threshold or more at "crimping", and the process corresponding to the time position is the abnormal process. It can be seen that it can be estimated.

  On the other hand, in the verification example according to the comparative example, as shown in FIG. 5B, from the start of the “adhesive application” to the end of the “crimping” is referred to as “first step + second step + third step”. However, elapsed time between processes was excluded. That is, this corresponds to the case where each process shown in FIG. 1C is to be learned, but one correlation model 12 d ′ is generated by batch learning without considering the elapsed time between the processes.

  In this case, as shown in FIG. 5B, according to “when normal quality component is input”, there are a plurality of time positions where the degree of deviation from the normal state is equal to or greater than the predetermined threshold, and the correlation model 12d ′ is in the normal state. It can be seen that manufacturing line 2 can not be modeled correctly.

  Further, according to the verification result of “at the time of inputting the abnormal quality”, since there are a plurality of time positions where the deviation degree from the normal state is equal to or more than the predetermined threshold, whether the process corresponding to the time position is an abnormal process It turns out that it is unknown.

  Next, the processing procedure executed by the abnormal process estimation device 10 will be described using FIG. FIG. 6 is a flowchart showing a processing procedure executed by the abnormal process estimation device 10.

  As shown in FIG. 6, the preprocessing unit 11b executes preprocessing based on the collection log 12a collected by the collection unit 11a (step S101). Then, the generation unit 11d executes generation processing based on the divided learning data set 12c extracted by the extraction unit 11c from the batch log group 12b preprocessed by the preprocessing unit 11b (step S102).

  Then, the evaluation unit 11e inputs the evaluation data set 12e extracted by the extraction unit 11c to each of the correlation models 12d-1, 12d-2, 12d-3 generated by the generation unit 11d, and executes the evaluation process (steps S103).

  Then, the determination unit 11f performs a determination process of estimating an abnormal process based on the evaluation result of the evaluation unit 11e (step S104). Although not shown, the notification unit 11g performs notification processing as needed based on the determination result of the determination unit 11f. Then, the process ends.

  The abnormal process estimation device 10 according to the embodiment is realized by, for example, a computer 60 configured as shown in FIG. 7. FIG. 7 is a hardware configuration diagram showing an example of a computer that realizes the function of the abnormal process estimation device 10. The computer 60 includes a central processing unit (CPU) 61, a random access memory (RAM) 62, a read only memory (ROM) 63, a hard disk drive (HDD) 64, a communication interface (I / F) 65, and an input / output interface (I). / F) 66, and a media interface (I / F) 67.

  The CPU 61 operates based on a program stored in the ROM 63 or the HDD 64 to control each part. The ROM 63 stores a boot program executed by the CPU 61 when the computer 60 starts up, a program depending on the hardware of the computer 60, and the like.

  The HDD 64 stores a program executed by the CPU 61, data used by the program, and the like. The communication interface 65 corresponds to a communication unit (not shown) with the manufacturing apparatus 100, receives data from another device through the communication network, sends it to the CPU 61, and transmits the data generated by the CPU 61 through the communication network. Send to another device.

  The CPU 61 controls an output device such as a display or a printer and an input device such as a keyboard or a mouse via the input / output interface 66. The CPU 61 acquires data from the input device via the input / output interface 66. Also, the CPU 61 outputs the generated data to the output device via the input / output interface 66.

  The media interface 67 reads a program or data stored in the recording medium 68 and provides the CPU 61 via the RAM 62. The CPU 61 loads the program from the recording medium 68 via the media interface 67 onto the RAM 62 and executes the loaded program. The recording medium 68 is, for example, an optical recording medium such as a digital versatile disc (DVD) or a phase change rewritable disc (PD), a magneto-optical recording medium such as a magneto-optical disk (MO), a tape medium, a magnetic recording medium, or a semiconductor memory. Etc.

  When the computer 60 functions as the abnormal process estimation device 10, the CPU 61 of the computer 60 executes the program loaded on the RAM 62 to obtain the collection unit 11a, the preprocessing unit 11b, the extraction unit 11c, the generation unit 11d, and the evaluation. The functions of the unit 11e, the determination unit 11f, and the notification unit 11g are realized. The HDD 64 realizes the function of the storage unit 12 and stores the collection log 12 a and the like.

  The CPU 61 of the computer 60 reads these programs from the recording medium 68 and executes them, but as another example, they may obtain these programs from another device via a communication network.

  As described above, the abnormal process estimation device 10 according to the embodiment includes the collection unit 11a, the generation unit 11d, the evaluation unit 11e, and the determination unit 11f. The collection unit 11a includes a plurality of sensors S-1 to S-n, and collects sensor data of the sensors S-1 to S-n in a manufacturing line 2 including a plurality of steps.

  The generation unit 11 d performs machine learning based on sensor data of normal quality while the manufacturing line 2 is in a normal state, by performing machine learning so that the elapsed time between the above steps is included in the feature vector. A plurality of correlation models 12d-1 to 12d-3 between the corresponding sensors S-1 to S-n are generated.

  The evaluation unit 11e inputs sensor data at any time of evaluation and evaluations for each of the steps, taking into account the elapsed time between the steps with respect to the sensor data, to the corresponding correlation models 12d-1 to 12d-3. The degree of deviation from the normal state in each of the steps is evaluated based on the output values of the correlation models 12d-1 to 12d-3 obtained thereby. Based on the degree of deviation, the determination unit 11 f determines an abnormal process that causes a quality abnormality of the product among the processes described above.

  Therefore, according to the abnormal process estimation device 10 according to the present embodiment, it is possible to accurately estimate an abnormal process which is a cause of product quality abnormality.

(Other embodiments)
By the way, in the embodiment described above, divided learning is performed in which the sensor data in each process and the elapsed time between processes are included in a vector, and the correlation model group 12d generated as a result separates from the normal state in the manufacturing line 2 An example of evaluating the degree has been given, but other parameters considered to contribute to the evaluation, for example, weather fluctuation data indicating the peripheral situation of the manufacturing line 2 may be further included in the vector.

  Further, in the embodiment described above, the case where the manufacturing line 2 bonds the support substrate W2 to the target substrate W1 is described as an example in the description using FIG. 5A and FIG. 5B. It does not limit the aspect of. For example, it may be a production line 2 in which the supplied parts are attached to the product body while applying an adhesive or the like to assemble the product.

  Further, in the above-described embodiment, deep learning is used as the machine learning algorithm, but the algorithm to be used is not limited. Therefore, machine learning may be performed by a regression analysis method such as support vector regression using a pattern discriminator such as SVM (Support Vector Machine) to generate the sensor correlation model group 12d. Here, the pattern discriminator is not limited to the SVM, but may be, for example, AdaBoost. Also, a random forest or the like may be used.

  Further effects and modifications can be easily derived by those skilled in the art. Thus, the broader aspects of the invention are not limited to the specific details and representative embodiments represented and described above. Accordingly, various modifications may be made without departing from the spirit or scope of the general inventive concept as defined by the appended claims and their equivalents.

DESCRIPTION OF SYMBOLS 1 abnormal process estimation system 2 manufacturing line 10 abnormal process estimation apparatus 11 control part 11a collection part 11b pre-processing part 11c extraction part 11d evaluation part 11e evaluation part 11f determination part 11g notification part 12 storage part 12a collection log 12b collective log group 12c division Learning data set 12d Correlation model group 12d-1 First correlation model 12d-2 Second correlation model 12d-3 Third correlation model 12e Evaluation data set 12f Evaluation information 12fa Deviation degree 12fb Contribution order 100 Manufacturing equipment S-1 to S- n sensor

Claims (7)

A collection step of collecting sensor data of the sensors in a production line comprising a plurality of sensors and comprising a plurality of steps;
The sensor corresponding to each of the steps is performed by executing machine learning so that the elapsed time between the steps is included in the feature vector based on the sensor data of the normal quality while the manufacturing line is in the normal state. Generating a plurality of correlation models between
An output value of the correlation model obtained by inputting the sensor data at the time of any evaluation and an evaluation component for each process in which an elapsed time between the processes is added to the sensor data to the corresponding correlation model. Evaluating the deviation from the normal state in each of the steps based on
A determination process of determining an abnormal process that is a cause of a product quality abnormality among the processes based on the degree of deviation. In the determination step,
The abnormal process estimation method according to claim 1, wherein the abnormal process is determined based on a time position at which the deviation degree is equal to or more than a predetermined threshold value. The method further includes a pretreatment step of performing pretreatment such that an elapsed time between the steps is taken into consideration.
The collecting step
Collecting the sensor data as log data divided for each process;
The pretreatment step is
A series of manufacturing processes in the manufacturing line by putting together the log data for each process into one while inserting a number of rows corresponding to the elapsed time between the processes between the log data for each process The abnormal process estimation method according to claim 1 or 2, wherein a batch log having one unit as one unit is generated. The pretreatment step is
The abnormal process estimation method according to claim 3, wherein a place corresponding to the sensor data is filled with 0 for a number of rows corresponding to an elapsed time between the processes. The generation step is
The abnormal process estimation method according to claim 3 or 4, wherein the machine learning is performed based on the collective log. A collection unit which has a plurality of sensors and collects sensor data of the sensors in a manufacturing line consisting of a plurality of processes;
The sensor corresponding to each of the steps is performed by executing machine learning so that the elapsed time between the steps is included in the feature vector based on the sensor data of the normal quality while the manufacturing line is in the normal state. A generation unit that generates a plurality of correlation models between
An output value of the correlation model obtained by inputting the sensor data at the time of any evaluation and an evaluation component for each process in which an elapsed time between the processes is added to the sensor data to the corresponding correlation model. An evaluation unit that evaluates the degree of deviation from the normal state in each of the steps based on
And a determination unit that determines an abnormal process that is a factor of product quality abnormality among the processes based on the degree of deviation. A collection procedure for collecting sensor data of the sensors in a production line comprising a plurality of sensors and comprising a plurality of processes;
The sensor corresponding to each of the steps is performed by executing machine learning so that the elapsed time between the steps is included in the feature vector based on the sensor data of the normal quality while the manufacturing line is in the normal state. Generation procedure that generates multiple correlation models between
An output value of the correlation model obtained by inputting the sensor data at the time of any evaluation and an evaluation component for each process in which an elapsed time between the processes is added to the sensor data to the corresponding correlation model. An evaluation procedure for evaluating the degree of deviation from the normal state in each of the steps based on
And a determination step of determining an abnormal step which is a factor of product quality abnormality among the steps based on the degree of deviation.