JP2008117381A - Time sequential data analyzer, time sequential data analysis system, time sequential data analysis method, program and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a time sequential data analyzer capable of performing optimum modeling. <P>SOLUTION: The time sequential data analyzer 2 for analyzing the time sequential data of a prediction object characteristic value in a prescribed system on the basis of the detected result information of two or more kinds of characteristic values in the system comprises: a target variable acquisition part 12 for acquiring the time sequential data composed of the past detected result of the prediction object characteristic value; a smoothed data extraction part 32 for calculating smoothed data for which the time sequential data are smoothed by a prescribed reference; a differential data extraction part 34 for calculating a difference between the time sequential data and the smoothed data as differential data; a smoothed data modeling part 42 for specifying the correspondence relation of time lapse and the variation of the smoothed data as a smoothed data variation model by analyzing the past variation of the smoothed data; and a differential data modeling part 44 for specifying the kind of the characteristic value affecting the variation of the differential data as a variation factor characteristic value and specifying the correspondence relation of the variation of the variation factor characteristic value and the variation of the differential data as a differential data variation model. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a time series data analysis apparatus for analyzing time series data.

  Conventionally, control charts are generally used for monitoring quality characteristics in product production lines and estimating causes in the event of an abnormality. The control chart is devised for managing the process by judging the variation in the characteristic value due to the accidental cause in the process and the variation in the characteristic value due to the abnormality cause. The control chart shows the characteristic value on the vertical axis, the time on the horizontal axis, the control line that is the characteristic value plotted in the horizontal direction along the time series, the upper limit of the allowable range as the product of the plotted characteristic value, and It consists of two upper and lower control limit lines plotted at positions determined by reasonably determining the lower limit.

  When characteristic values representing the state of a process are plotted as points, if all the points are within the upper and lower two control limit lines and there is no characteristic tendency in the arrangement of points, the process is in a stable state. Can be considered. On the other hand, if the point is outside the control limit line or if a characteristic tendency appears in the arrangement of the points, the process can be regarded as not in a stable state. In this case, it is necessary to determine that an abnormal state has occurred in the process, investigate the cause, and take action.

  Because quality characteristics vary due to various factors that make up the production line, including factors that cannot be measured, the control chart determines variations in characteristic values due to accidental causes in the process and variations in characteristic values due to abnormal causes However, it is widely used to check whether the process is in a stable state.

  At present, in a production line that manufactures products through a plurality of manufacturing processes, final quality evaluation is performed after passing through all the processes. That is, quality evaluation is not performed in the middle process. For this reason, it is difficult to make adjustments and corrections for defective products caused by the initial process at the stage of quality evaluation.

  In order to solve the above problems, various methods have been devised for predicting the characteristic values obtained by the final quality evaluation before performing the final quality evaluation.

  Patent Document 1 discloses a time-series data prediction apparatus for the purpose of improving the prediction performance of a time-series data prediction apparatus.

  In the above prediction apparatus, the first time series prediction means constructed using the feedforward type neural network, the second time series prediction means constructed using the ARIMA model, the use conditions, the first and the second Input the prediction results and actual measurement results of 2 time series prediction means, and compare the prediction results and actual measurement results for each usage condition, and measure the superiority of the prediction capability according to the usage conditions for each time series prediction means Use characteristic measurement means, use conditions, prediction results, and measurement results from the use characteristic measurement means are input, and based on the use conditions and measurement results, the weighted addition method is used to integrate the prediction results for final prediction. It is shown that an integrated judgment means for outputting the result is provided.

  Patent Document 2 discloses a prediction method using a recurrent network.

  In the above prediction method, prediction is performed by a recurrent network having feedback coupling and learning time series data. Specifically, the network is initialized, the network is predicted up to the continuous point of the target output data as preprocessing, the network state at the end of the prediction is stored, and the above-mentioned storing step is performed. It is shown that the method includes a step of performing prediction by the network from the state of the generated network.

  Further, Patent Document 3 captures a process leading to a dangerous work in a work as a chain of time-series work procedures and a change in work environment, analyzes the causal relationship in time series, and predicts a dangerous work. A prediction system is disclosed.

  In the above prediction system, the work procedure to be analyzed is decomposed and the context is organized, the factors of the disaster are extracted, the causal relationships are organized, the work and the disaster factors are combined and expressed in a Petri net, It is shown that a simulation relating to work is performed using a Petri net, the work method and work environment are improved, and the process is repeated until a predetermined safety is satisfied.

  Further, Patent Document 4 discloses a method for predicting time-series data that enables highly accurate prediction from time-series data of a short time in the past.

In the above prediction method, not only the time-series data to be predicted but also other correlated time-series data are embedded, and these are combined to create an expanded vector, which is localized in the vector space. It is shown that a predicted value is obtained by performing fitting or the like.
JP 7-049883 A (published on March 31, 1995) JP 10-283335 A (published October 21, 1998) Japanese Patent Laid-Open No. 2001-351057 (released on December 21, 2001) JP 9-282306 A (published October 31, 1997)

  Control charts generally used for monitoring quality characteristics can show trends in quality, but whether the cause is in the equipment on the production line or the surrounding environment I can't judge.

  Further, as described in Patent Documents 1 and 2, prediction of collection quality characteristics using smoothing variation models such as AR, ARMA, and ARIMA, and differential variation models such as NN, recurrent NN, and RBF network In the prediction of the final quality characteristic using the method and the prediction combining the above two methods, attention is paid only to the change in the quality characteristic of the final process, and thus it is not possible to cope with the sudden change. Furthermore, when the factors of the previous process that affect the quality characteristics of the final process vary, there is a problem that the quality characteristics of the final process also vary and prediction becomes difficult.

  In addition, as described in Patent Document 3, in the configuration for modeling the various characteristics of the production line using the statistics and the final quality characteristics, and for structuring the relationship between the various characteristics of the production line and the final quality characteristics, In an actual production line, various factors including the final process are directly or indirectly affected, and change due to factors that cannot be measured. Therefore, it is difficult to construct an accurate causal relationship. In particular, a change due to an unmeasurable factor causes a problem of affecting modeling as an error.

  Further, as described in Patent Document 4, the final quality characteristics are approximated by time series prediction, and the residual of the trend and the actual measurement is influenced by other measurable factors. In a configuration where the factor and the residual are statistically correlated, and the final prediction is performed by adding the approximate approximation of the former and the statistical prediction of the latter, unpredictable factors are mixed in the time-series prediction. Since the accuracy is lowered, there arises a problem that the reliability of the statistically associated data is lowered.

  The present invention has been made in view of the above problems, and its purpose is to perform modeling that is as optimal as possible by extracting fluctuations that have been determined to have a correlation with a predicted characteristic value. It is to realize a time-series data analysis apparatus capable of performing the above-described process.

  In order to solve the above-described problem, the characteristic value prediction apparatus according to the present invention analyzes time series data of prediction target characteristic values in the system based on detection result information of multiple types of characteristic values in a predetermined system. A data analysis device that calculates time-series data acquisition means for acquiring time-series data consisting of past detection results of the prediction target characteristic value, and smoothed data obtained by smoothing the time-series data according to a predetermined reference Smoothing data acquisition means, difference data acquisition means for calculating the difference between the time series data and the smoothed data as difference data, and analysis of past fluctuations in the smoothed data, thereby allowing the passage of time and the smoothing data. Smoothing data modeling means for specifying the correspondence relationship with the fluctuation of the smoothed data as a smoothing data fluctuation model, and the above-mentioned affecting the fluctuation of the difference data It identifies the type of sexual values as variables characteristic value, and a differential data modeling means for identifying a correspondence between the variation of the change and the difference data of the fluctuation factor characteristic value as difference data variation model.

  According to the above configuration, by acquiring time-series data consisting of past detection results and smoothing them according to a predetermined standard to obtain smoothed data, changes in the environment, the passage of time, and other unexpected Variations caused by factors can be averaged and smoothed. This smoothed data can be considered to show the original characteristic value from which the variation is removed. Therefore, it is considered that the magnitude of the variation can be extracted by taking the difference between the characteristic value obtained as the smoothed data and the measured value.

  Therefore, by modeling these smoothed data and difference data, variation in the characteristic value due to changes in the environment, the passage of time, etc., so-called temporal changes and the external environment, and sudden factors, for example, Data indicating a specific causal relationship can be separated into variations in characteristic values caused by events that can be obtained from sensor information or the like. Accordingly, for example, a modeling method suitable for creating a smoothing variation model, a difference variation model, or the like can be applied to each characteristic value according to the type of variation in the characteristic value.

  By each of the above steps, a variation factor that varies the prediction target characteristic value is separated according to the type, and a model to which an appropriate modeling method according to the type of the variation factor is applied can be created.

  In the time-series data analysis device, a radial basis function network having an intermediate layer neuron is used as the difference data modeling means, and the difference data modeling means includes the fluctuation of the variation factor characteristic value and the difference data. Learning data generating means for generating learning data including fluctuations, first center point selecting means for selecting the center point of the intermediate layer neuron based on the target value of the difference data from the learning data, and based on the selected center point The network construction means for constructing the radial basis function network and the network construction means are constructed by calculating the output value of the intermediate layer and calculating the weight value of the neuron based on the calculated output value of the intermediate layer. Of the difference data by applying the above learning data to the measured radial basis function network And a second center point selecting means for selecting a new center point of the intermediate layer neuron based on an error between the calculated estimated value and the fluctuation of the difference data. The output value of the intermediate layer is calculated based on the center point selected by the first and second center point selection means, and the weight value of the neuron is calculated based on the calculated output value of the intermediate layer.

  According to the above configuration, it is possible to extract a point to be improved in prediction accuracy and improve the prediction accuracy of a desired point, so that a measure is implemented by accurately predicting a point to be finally determined. Alternatively, it is possible to accurately determine that the operation is not performed.

  As a specific example, for example, in the case of predicting the final quality in the production line, by accurately predicting whether or not the final quality is good by accurately predicting the value near the control limit value, the countermeasure can be taken. Since it can be accurately determined whether or not to implement, the cost of measures for the product can be optimized.

  As another specific example, for example, in an office or factory, if the power consumed in the future is predicted and the demand contracted with the power supplier may be exceeded, the operation of the device that consumes the power In the case of a system that takes countermeasures such as temporarily stopping the system, it is possible to accurately determine whether the countermeasure is implemented or not by accurately predicting a value around the preset demand power. It is possible to minimize the impact on the business due to excess demand.

  In the time-series data analysis device, a radial basis function network having an intermediate layer neuron is used as the difference data modeling means, and the difference data modeling means includes the fluctuation of the variation factor characteristic value and the difference data. Learning data generating means for generating learning data including fluctuations, representative point selecting means for selecting a representative point matching the intermediate layer neuron from the learning data by the k-means method, and an intermediate layer based on the selected center point Network construction means for constructing the radial basis function network by calculating the weight value of the neuron based on the calculated output value of the intermediate layer.

  According to said structure, since the intermediate | middle layer neuron made into a representative point is selected by k-means method, there exists an effect that the learning of a radial basis function network can be performed in a short time.

  The time-series data analysis apparatus further includes a fluctuation factor estimating unit that estimates a fluctuation factor of the characteristic value at the prediction target time point from the smoothed data fluctuation model and the difference data fluctuation model.

  According to the above configuration, from the smoothed data variation model and the difference data variation model, whether the variation of the target characteristic value is a variation due to a time series change, a variation due to a time series change, or a sudden variation, for example, The cause of the fluctuation can be identified from a certain measurement result measured by the sensor, it is judged whether the causal relation is a clear fluctuation or the causal relation is a clear fluctuation, and the fluctuation factor of the characteristic value can be estimated.

  The time-series data analysis apparatus further includes a prediction unit that predicts a characteristic value at a prediction target time point from the smoothed data fluctuation model or the difference data fluctuation model.

  According to said structure, the fluctuation | variation of the target characteristic value in a certain time can be estimated from the said smoothing data fluctuation | variation model and the said difference data fluctuation | variation model.

  In the time-series data analysis device, a difference data prediction unit that acquires a characteristic value of the variation factor at the time of the prediction target and calculates a predicted value of the difference data based on the difference data variation model; The smoothed data predicting means for calculating the predicted value of the smoothed data at the time to be predicted based on the smoothed data fluctuation model, the prediction result by the difference data predicting means, and the smoothed data predicting means And a prediction integration unit that calculates a predicted value of the measurement target characteristic value based on a prediction result.

  According to the above configuration, more accurate prediction can be performed by performing prediction using a plurality of models generated by an optimal modeling method and adding the prediction results based on the plurality of prediction models. it can.

  In the time series data analysis apparatus, the plurality of types of characteristic values in the predetermined system are sensor information detected in each step of the production line in the production system, and the time series data to be predicted is It is characterized by time-series data of quality characteristics of production objects in the production line.

  According to the above configuration, the prediction target time point is predicted at a stage before the prediction target time point, for example, at a stage where a product produced on the production line can be collected and an intermediate process can be performed again. Adjustments can be made before a defective product is created to prevent the generation of unexpected defective products.

  In the time-series data analysis apparatus, the temporal abnormality cause estimation reference value, which is a threshold for managing the change in quality characteristics accompanying the temporal change of the production object, is obtained from the prediction result by the smoothed data prediction unit. Threshold setting means for setting an upper limit value and a lower limit value is further provided.

  In the time-series data analysis apparatus, the upper limit of the sudden abnormality cause estimation reference value, which is a threshold value for managing the quality characteristic abnormality accompanying the sudden change of the production object, based on the prediction result by the difference data prediction unit. Threshold setting means for setting the value and the lower limit value is further provided.

  According to the above configuration, the judgment criteria for abnormality are set for each of the change in quality characteristics over time and the abnormality in quality characteristics due to sudden change with respect to the control limit set in the straight line. Can do.

  In addition, the time-series data analysis and the objective variable inspection device for acquiring sensor information from the sensor and inputting it as time-series data to the time-series data acquisition means, the sensor information from the sensor, and the smoothing A model creation system can be configured by including an explanatory variable inspection apparatus for inputting data as learning parameters to the data modeling means and the differential data modeling means, and a management apparatus for managing acquisition of the sensor information.

  The determination method of the present invention is a model generation method for detecting a plurality of types of characteristic values in a predetermined system and creating a model that can identify a variation factor of a prediction target characteristic value in the system based on detection result information. Obtaining time series data comprising past detection results of the prediction target characteristic value; calculating smoothed data obtained by smoothing the time series data according to a predetermined reference; and the time series data and the smoothing Calculating the difference between the smoothed data as differential data and analyzing the past fluctuations in the smoothed data, thereby specifying the correspondence between the passage of time and the smoothed data as a smoothed data fluctuation model Identifying the step and the type of the characteristic value that affects the fluctuation of the difference data as the fluctuation factor characteristic value, and It is characterized in that it comprises a step of identifying correspondence between the variation of the difference data as the difference data variation model.

  According to said structure, there can exist an effect similar to the above-mentioned.

  In addition, the said determination method can be performed on a computer by control of a computer. Furthermore, by storing the above program in a computer-readable recording medium, the program can be executed on any computer.

  The characteristic value prediction apparatus and characteristic value prediction method according to the present invention, as described above, correlate with the characteristic value by removing the smoothed objective variable from the objective variable based on a plurality of sensor information acquired in the past. The cause-specific variation determined to exist can be extracted, so a prediction model that eliminates the cause-specific variation and a prediction model based on the cause variation are generated, a model is created by applying the optimal method, and the characteristic value There is an effect that it is possible to predict.

Embodiment 1
An embodiment of the present invention will be described below with reference to FIGS. In the following, quality characteristic prediction in a substrate mounting process from mounting of a component of a substrate to welding of a solder will be described as an example, but the present invention is not limited to this. For example, it may be a power consumption prediction based on the temperature of each place. Any prediction can be used as long as appropriate objective variables and setting variables are defined, different characteristics are separated, and modeling is performed for each characteristic.

  In the above configuration, the operation when the prediction system of the present invention predicts the quality characteristic will be described with reference to FIG. FIG. 1 is a functional block diagram illustrating an outline of a prediction system using the prediction device according to the present embodiment. The prediction system 1 of this embodiment includes an acquisition unit 10, a learning unit 20, a prediction unit 70, and a process processing device 80. Details of each part will be described below.

  The acquisition unit 10 acquires data such as sensor information from the inspection device in the process processing apparatus 80 and inputs the data to the learning unit 20. The acquisition unit 10 includes an objective variable acquisition unit 12 and an explanatory variable acquisition unit 14. The objective variable acquisition unit 12 acquires data corresponding to the objective variable from the inspection device in the process processing device 80, and the explanatory variable acquisition unit 14 acquires data corresponding to the explanatory variable from the inspection device in the process processing device 80. To do. Details of the objective variable, the explanatory variable, and the data corresponding to each variable will be described later.

  The learning unit 20 creates the difference variation model 52 and the smoothing variation model 54 based on the data input from the acquisition unit 10. The details of how to create each model will be described later. The learning unit 20 includes a characteristic separation unit 30, a prediction learning unit 40, a difference variation model 52, a smoothing variation model 54, and a parameter setting unit 60.

  The characteristic separation unit 30 extracts characteristic values from the input target variable data and uses them for predictive learning. The characteristic separation unit 30 includes a smoothed data extraction unit 32 and a difference data extraction unit 34.

  The smoothed data extraction unit 32 extracts characteristic values for trend fluctuations, that is, fluctuations that cannot be expressed by causal relationships. As specific examples, it refers to changes in solder over time in board mounting, changes in ambient temperature, and the like. The characteristic value of the trend variation extracted by the smoothed data extraction unit 32 is used to create a smoothing variation model.

  The difference data extraction unit 34 extracts a characteristic value with respect to variation due to the causal relationship. Specific examples include vertical displacement, lateral displacement, excessive amount of solder, and other direct factors in board mounting. The characteristic value of the causal relationship variation extracted by the difference data extraction unit 34 is used to create a difference variation model.

  Here, the term of the causal relationship in this embodiment is demonstrated. The causal relationship here refers to a change in the characteristic value based on the substrate itself used for measurement or the component itself to be measured. It is assumed that there is no cause in the board or component, and the characteristic value that varies mainly depending on the device of the production line and the surrounding environment cannot be expressed by the causal relationship. Solder aging means that the solder surface is kept in contact with air for a long time when the solder is stored in the production line and printed on the board when necessary and used for joining parts. It means that the viscosity of solder changes.

  The prediction learning unit 40 creates each model based on each characteristic value extracted by each unit of the characteristic separation unit 30. The prediction learning unit 40 includes a smoothed data modeling unit 42 and a difference data modeling unit 44. The smoothed data modeling unit 42 creates a smoothed variation model 54 based on the data extracted by the smoothed data extracting unit 32. The difference data modeling unit 44 creates the difference variation model 52 based on the data extracted by the difference data extraction unit 34.

  In this embodiment, as an example, AR is used for the smoothing fluctuation prediction modeling function, and a radial basis function network (RBFN), which is a kind of hierarchical neural network, is used for the difference data modeling function. This is not a limitation. For example, autoregressive moving average (ARMA), autoregressive integrated moving average (ARIMA), or the like may be used for smoothing fluctuation prediction, and neural network (NN), recurrent NN, or the like may be used for differential fluctuation prediction. It may be used.

  The difference variation model 52 is created by the difference data modeling unit 44 and used for smoothing variation prediction. The smoothed fluctuation model 54 is created by the smoothed data modeling unit 42 and used for difference fluctuation prediction. The parameter setting unit 60 adjusts the parameter of the smoothing / smoothing data modeling function of the difference data modeling 44 and the parameter of the characteristic separation function of the characteristic separation unit 30 based on the difference fluctuation prediction error.

  The prediction unit 70 is a functional block that actually predicts a final quality characteristic using a result learned using each created model. The prediction unit 70 includes a difference variation prediction unit 72, a smoothing prediction unit 74, a prediction integration unit 76, and a cause estimation unit 78. The difference fluctuation prediction unit 72 performs fluctuation prediction based on the causal relationship based on the difference fluctuation model 52. The smoothing prediction unit 74 performs fluctuation prediction based on trend fluctuation based on the smoothing fluctuation model 54. These two prediction results are integrated in the prediction integration unit 76, and final quality characteristics are predicted.

  In the present invention, the measurement value in the previous process is defined as an explanatory variable, and the final quality characteristic prediction is defined as an objective variable. In the board mounting example of this embodiment, the vertical shift of the component, the horizontal shift of the component, etc. are used as the explanatory variables, and the fillet length is used as the objective variable which is the final quality characteristic, but is not limited to this. Absent.

  The cause estimating unit 78 can detect a sign of abnormality using the predicted value of the final quality characteristic integrated by the prediction integrating unit 76 and the value of the control limit, and investigate the cause. In addition, the control limit value is dynamically changed, the cause based on the time series is fed back to the manufacturing process for improvement, and the final process in board mounting is the reflow process that heats the solder and fixes the parts Before performing the steps, it is possible to remove the substrate that is likely to be defective and improve the substrate.

  The process processing apparatus 80 is an apparatus for actually measuring each work in the substrate mounting and characteristic values in each work stage. The process processing device 80 includes a print inspection device 82, a displacement inspection device 84, a fillet length inspection device 86, a process management device 90, a printing device 92, a mounting device 94, and a reflow device 96.

  The print inspection device 82 is a sensor device for inspecting the amount of cream solder printed on the substrate. The deviation inspection device 84 is a sensor device for inspecting the positional deviation of the electronic components arranged on the substrate. The fillet length inspection device 86 is a sensor device for inspecting the fillet length of solder fixed after the reflow process. Each sensor device may be an image input device such as an optical camera, or may be a test device that performs a continuity test or a component failure test.

  This is a control device for controlling and monitoring the process management device 90, each inspection device, and each process processing device. In the present embodiment, the function is realized by a computer (PC). The process management device 90 may stop the process processing or correct the processing operation based on the prediction result of the prediction unit 72, the inspection result in each inspection device, or the like.

  Based on the instructions of the printing device 92 and the process management device 90, a cream solder printing process is performed on the printed wiring board. Based on instructions from the mounting device 94 and the process management device 90, the electronic components are arranged on the board. The reflow device 96 applies heat to the substrate on which the components and solder are arranged based on an instruction from the process management device 90, melts the cream solder, and joins the components to the substrate.

  Next, the flow of processing for actually performing prediction using the prediction system 1 of the present embodiment will be described with reference to FIG. Processing performed in the prediction system 1 is roughly divided into a learning process and a prediction / determination process. First, the processing procedure will be described from the learning process.

  In the learning process, learning is performed by creating a prediction model based on input data such as sensor information acquired from the system. First, the acquisition unit 10 of the prediction system 1 acquires objective variables and explanatory variables from the system. In this embodiment, the objective variable is described as the fillet length of the component A in board mounting, and the explanatory variable is described as the lateral shift of the component A, the vertical shift of the component A, and the solder amount of the component A, but is not limited thereto.

  First, explanatory variables will be described. After completion of each process of board mounting, inspection is performed by an inspection apparatus in order to determine whether the processing performed in each process is acceptable. As a specific example, after the process of printing the printed wiring by the printing device 92 is completed, the print inspection device 82 acquires data for judging whether printing is good or bad. In addition, after the process of mounting the component by the mounting device 94 is completed, data is collected by the deviation inspection device 84 in order to determine whether the component mounting position is acceptable. In the present embodiment, the lateral displacement of the component A, the longitudinal displacement of the component A, and the amount of solder of the component A are inspected by an inspection apparatus and acquired as explanatory variables.

  Next, the objective variable will be described. In the present embodiment, parameters are set as characteristic values for finally predicting the fillet length of the part A. Fillet refers to the gentle curved surface formed between the base or land and the component or terminal by solder. Fillet length refers to solder characteristics such as component misalignment, amount of solder, wetting performance, cream solder Varies with prescription, land design, heating conditions, and other factors. The fillet length is determined by applying heat to components and solder placed on the substrate, cooling the substrate, and completing the reflow process.

  In the substrate mounting process, the fillet length is inspected by the fillet length inspection device 86 for the mounting substrate for which the reflow processing has been completed by the reflow device 96. The fillet length data obtained here is acquired by the objective variable acquisition unit 12 and input to the characteristic separation unit 30.

  The characteristic separation unit 30 smoothes the acquired objective variable to obtain an average value of data in a predetermined continuous period corresponding to the most recent times, and the objective variable is in accordance with the flow of time. Let it be a continuous smooth graph. This is because in the actual objective variable, there is a possibility that the objective variable is fluctuated due to an unexpected factor, for example, a deviation in the arrangement of parts.

  The variation of the objective function will be described with reference to FIG. FIG. 13 is a graph showing changes in values of the objective variable and the explanatory variable along the flow of time.

  The graph 101 in the uppermost diagram shows changes in the objective variable (quality characteristics). The vertical axis represents the fillet length of the part A. The graph 102 in the second diagram shows the change in the explanatory variable (vertical deviation of the part A). The vertical axis represents the component shift. A graph 103 in the third diagram shows a change in the explanatory variable (lateral deviation of the part A). The vertical axis represents the deviation of the part A. The input of these explanatory variables is processed in order from the process corresponding to the lower graph in the production line.

  The horizontal axis of the three graphs indicates the elapsed time. A dotted line graph 100 in the top graph is a smoothed value of the graph 101. The graph 100 obtained by performing the smoothing process is a surrounding environment that does not originate from the substrate itself, such as temperature, humidity, machine condition in the manufacturing process, solder deterioration, etc., excluding sudden factors. Due to the above, it can be considered that the fluctuation is associated with the elapsed time. In this embodiment, such variation is called “time-series variation (trend variation, time-dependent variation)”, and it does not vary with the elapsed time due to the placement of the substrate and the parts placed on the substrate. Such fluctuation is called “causal fluctuation”.

  In the characteristic separation process, smoothed data is calculated by smoothing, and a process of extracting the residual between the smoothed data and the actually measured value as difference data is performed. The smoothing process and the characteristic separation process performed by the characteristic separation unit 30 will be described later in detail. By separating the smoothed data and difference data by the above processing, modeling can be performed using a modeling technique suitable for each variation. Returning again to FIG. 2, the description of the learning process will be continued.

  The objective variables separated into the smoothed data and the difference data by the characteristic separating unit 30 are input to the smoothed data modeling unit 42 and the difference data modeling unit 44, respectively. The smoothed data modeling unit 42 calculates the smoothing variation model 54 based on the objective function classified into the smoothed data, that is, the smoothed objective variable and the input including feedback from the smoothing variation model 54. Generate. The difference data modeling unit 44 generates a difference variation model 52 based on the objective variable classified as difference data and a plurality of explanatory variables.

  In the smoothing variation model 54 and the difference variation model 52, the parameter setting unit 60 updates each prediction model based on the objective variable and the explanatory variable input one after another along the elapsed time. Based on this, the parameter of the model generation process in each modeling unit and the parameter of the characteristic separation process in the characteristic separation unit 30 are adjusted.

  Note that a method suitable for each prediction item is used to create each model. Here, model generation is performed using AR for smoothing fluctuation prediction modeling and an RBF network for differential fluctuation prediction modeling, but not limited to this, NN, recurrent NN, and statistics are used. Modeling various characteristics and final quality characteristics, structuring the relationship between various characteristics and final quality characteristics of the production line, approximating the final quality characteristics by smoothing variation and calculating the factors of the previous process and the residual A method of statistically associating and predicting by adding approximate prediction and statistical prediction, and various other methods may be used.

  Learning is performed by adjusting each setting parameter by repeating the above processing, and generating, updating, and constructing a prediction model.

  Next, the prediction / determination process will be described. In the prediction / judgment process, the value of the next target variable is predicted from a predetermined number of target variables in the past, and the current target variable value is predicted from several current explanatory variables in a series of steps. Specifically, in the example of FIG. 13, a3 is predicted from the values of the objective variables a1 and a2, and the value of the objective variable a3 is predicted from the values of the explanatory variables b3 and c3. In this example, the predetermined number is the two most recent objective variables.

  Note that, unlike the learning process, only some explanatory variables are input in the prediction / determination process. In the example of FIG. 2, it is assumed that only two of the three explanatory variables are input and only directly influenced variables that are found from the learning results of the learning process are used. In the present embodiment, only the lateral shift of the part A and the vertical shift of the part B are input as explanatory variables, and the solder amount of the part A is not input as a variable that does not directly affect the prediction.

  First, the trend change prediction based on the smoothing fluctuation prediction will be described. In the objective variable acquired from the system, the trend data is extracted in the smoothed data extraction unit 32 based on the parameters adjusted in the learning process. The extracted trend variation is input to the smoothing prediction unit 74. The smoothing prediction unit 74 calculates a predicted value of the trend variation based on the trend variation extracted by the smoothing data extraction unit and the difference variation model 54.

  Next, prediction of difference data based on difference fluctuation prediction will be described. As described above, the explanatory variables acquired from the system are input to the difference fluctuation prediction unit 72 except for the explanatory variables determined not to directly affect the prediction. The difference fluctuation prediction unit 72 calculates a predicted value of the difference data based on the input explanatory variable and the difference fluctuation model.

  The trend variation calculated based on the above input values and the predicted value of the difference data are integrated by the prediction integration unit, and the final predicted value is calculated. Since each prediction model is modeled using a method suitable for the model, it is possible to predict the final prediction value more accurately by integrating the prediction values of each model.

  From the final predicted value calculated in this way and the input explanatory variable, the cause estimating unit 78 determines whether the cause of the change in the quality characteristic is a causal variation such as a production line, a production line device, or an external environment. It is possible to determine whether there is a trend change.

  Next, details of processing performed in the learning process will be described with reference to FIG. FIG. 3 is a flowchart showing the flow of actual processing performed in the learning process.

  First, when the learning process is started, the explanatory variable group and the objective variable are sequentially measured from the mounting boards that flow one after another by the objective variable acquisition unit 12 and the explanatory variable acquisition unit 14 of the acquisition unit 10 (S101). , S102). The measured objective variable is smoothed by the smoothed data extraction unit 32 and extracted as smoothed data (trend fluctuation) (S103). Thereafter, the difference data extraction unit 34 extracts the residual between the extracted smoothed data and the objective variable measured in S101 as difference data.

  The smoothed data (trend variation) extracted in S103 is modeled as a smoothed variation model 54 by the smoothed data modeling unit 42 (S105). Further, the difference data extracted in S104 is modeled as a difference variation model 52 by the difference data modeling unit 44 (S106). Thereafter, the parameter setting unit 60 adjusts parameters to be reflected in the smoothed data modeling unit 42 and the characteristic separation unit 30 based on the data used for the modeling processing in S105 and S106 (S107). Details of the parameter adjustment processing will be described later. The result of parameter adjustment is reflected at the position (1) in the flowchart of FIG. 3, and is reflected in the extraction process in S103 and the modeling of the difference variation in S106.

  Next, details of the parameter adjustment processing will be described with reference to FIG. FIG. 4 is a flowchart showing the flow of parameter adjustment processing performed by the parameter setting unit 60.

  First, when the parameter adjustment process is started, the parameter setting unit 60 calls the objective variable and the explanatory variable group acquired by the objective variable acquisition unit 12 and the explanatory variable acquisition unit 14 (S201). Each called variable performs a prediction process using the prediction functions of the difference fluctuation prediction unit 72 and the smoothing prediction unit 74. Details of the prediction process will be described later.

  Next, the parameter setting unit 60 extracts the reproducibility of the causal variation of the explanatory variable from the prediction result calculated using the prediction function (S204). Based on the extracted reproducibility, the parameter setting unit 60 determines whether the reproducibility of the causal relationship is sufficient (S204). If it is determined that the reproducibility of the causal relationship is sufficient (Yes in S204), the parameter adjustment process ends. If it is determined that the reproducibility of the causal relationship is not sufficient (No in S204), the process proceeds to S205.

  In S205, it is determined whether the causal relationship is reproduced by being smoothed (S205). If it is determined that the causal relationship has been smoothed and reproduced (Yes in S206), the process proceeds to S206. If it is determined that the causal relationship has not been smoothed and reproduced (No in S206), the process proceeds to S207.

  In S206, parameter adjustment is performed to shorten the data size used for smoothing the objective variable. As a result, the parameter adjustment is reflected in (1) of FIG. Similarly, in S207, parameter adjustment is performed to increase the data size used for smoothing the objective variable. As a result, the parameter adjustment is reflected in (1) of FIG.

  Next, details of processing performed in the prediction / determination process will be described with reference to FIG. FIG. 5 is a flowchart showing an actual processing flow performed in the prediction / determination process.

  First, when the prediction / determination process is started, the explanatory variable group and the objective variable are sequentially measured from the mounting boards that flow one after another by the objective variable acquisition unit 12 and the explanatory variable acquisition unit 14 of the acquisition unit 10. (S301, S302).

  The measured objective variable is smoothed by the smoothed data extraction unit 32 and extracted as smoothed data (trend fluctuation) (S303). Thereafter, the smoothing prediction unit 74 performs smoothing fluctuation prediction based on the smoothed smoothed data and the smoothing fluctuation model 54 reflecting the learning process (S305). The result of the smoothed fluctuation prediction is input to the prediction integration unit 76.

  On the other hand, from the measured explanatory variable group, only explanatory variables that directly affect the objective variable are selected and input to the difference fluctuation prediction unit 72 (S304). Thereafter, the difference fluctuation prediction unit 72 performs difference fluctuation prediction based on the selected explanatory variable group and the difference fluctuation model 52 reflecting the learning process (S306).

  Thereafter, the prediction integration unit 76 adds the smoothed fluctuation prediction performed in S305 and the difference fluctuation prediction performed in S306 to calculate a final prediction result (S307). Thereafter, based on the calculated final prediction result, the cause estimating unit 78 performs defect cause estimation processing (S308). Details of the cause estimation process will be described later.

  Next, details of processing performed in the cause estimation processing will be described with reference to FIG. FIG. 6 is a flowchart showing the flow of actual processing performed by the cause estimating unit 78 in the cause estimating process.

  First, when the cause estimation process is started, the cause estimation unit 78 checks whether or not the final predicted value exceeds the management limit (S401), and determines (S402). If the final predicted value exceeds the management limit (Yes in S402), the process proceeds to S403. If the final predicted value does not exceed the management limit (No in S402), the process ends.

  In S403, it is confirmed whether the smoothed data prediction (smoothing fluctuation prediction) exceeds the control limit (S403). If the control limit is exceeded (Yes in S404), the process proceeds to S405. If the control limit is not exceeded (No in S404), the process proceeds to S406. If the control limit is exceeded, the main cause of the abnormality is a change with time, and processing for notifying this is performed (S405). If the control limit is not exceeded, the main cause of the abnormality is a factor in the previous process, and processing for notifying this is performed (S406). Thereafter, the cause estimation process is terminated.

  Next, details of processing performed in the differential data modeling unit will be described with reference to FIGS. FIG. 7 is a flowchart showing the flow of processing of a radial basis function network (RBFN) that performs differential data modeling. FIG. 8 is a diagram illustrating a table for explaining a list of variables used for various calculations used when processing is performed by the RBFN. FIG. 9 is a diagram illustrating the structure of the RBFN prediction model used in the differential data modeling unit. FIG. 10 is a diagram illustrating a Gaussian function used in the differential data modeling unit. FIG. 11A and FIG. 11B are diagrams showing an outline of the arrangement of default functions in the above RBFN prediction model. FIG. 12 is a diagram for explaining selection of the center point in the RBFN prediction model.

  Here, an outline of RBFN will be described. RBFN is a feed-forward type network consisting of three layers (see FIG. 9), and can approximate an arbitrary nonlinear function in the same way as a multilayer perceptron. RBF (radial basis function) refers to a function that reacts locally in the input space. That is, a function that takes a maximum value or a minimum value at the center and monotonously decreases or increases as the distance from the center increases. Although many such functions exist, a frequently used one is a Gaussian function as shown in FIG.

  The Gaussian function is expressed by the following mathematical formula 1.

Here, h _j (x) is the output value from the intermediate layer element, x (= x ₁ , Λ, x _n ) is the input data from the input layer element, c _j is the center point of the Gaussian function, and r is the Gaussian function Points to the radius.

  As shown in FIG. 9, the basis function network model has three layers: an input layer with n elements, an intermediate layer with m elements, and an output layer with l elements. Each layer is composed of elements that input and output. The elements between the layers are connected by wiring, but the elements in the same layer are not coupled to each other.

X1 of this model, ..., the _{x n,} a 1, _..., can be a predictive model by giving _{a t} + Delta] t as learning data in a t, 0 (x). However, a ₁ ,..., A _t , a _{t + 1} ,... A _t + Δt are time-series data, and Δt indicates a previous period to be predicted when the current time is t. Note that h (x) is a vector whose elements are x ₁ ,..., X _n , and 0 (x) is the same.

The output of each element in the intermediate layer uses a bell-shaped Gaussian function which is one of the basis functions (FIG. 10). In the predictive learning model using the basis function, this means that the weight (w _i ) from the intermediate layer to the output layer, the center position c _{j of the} default function h (x), and the radius r are determined in FIG. In RBFN, a nonlinear function is expressed by a weighted sum of basis functions. The shape of RBFN changes by changing the center, width, and coupling weight of the basis function. Based on data that is actually measured values, learning in the RBFN is to estimate an appropriate parameter so that an error between the actually measured value data and the output of the RBFN is small.

  FIG. 11 shows an image of the arrangement of basis functions. As shown in FIG. 11A, individual basis functions are arranged, and as shown in FIG. 11B, response functions are created.

The shape of RBFN changes by changing the center, width, and coupling weight of the basis function. That is,
Learning in the RBFN is to estimate an appropriate parameter based on the data so that an error between the data and the output of the RBFN is reduced.

  Next, the flow of processing in RBFN will be described. When the differential data modeling process is started, an explanatory variable is selected (S501). Specific contents of the explanatory variables will be described later. Next, learning data is created (S502). In the present embodiment, as the learning data, the past measured values and the causal variation are used as a set of RBFN input / output relationships for the effective explanatory variables.

Thereafter, in S503, center points that match the number of intermediate layer neurons are selected from the set created in S502 using the k-means method. In S504, the representative point selected in S503 is positioned as the center point of each neuron in the intermediate layer, and the output values of the intermediate layer are calculated for all explanatory variable data sets according to the RBF having the base radius r. Finally, in S505, a net weight w _i between the intermediate layer and the output layer is calculated. At this time, the normalization parameter ε is used so that the calculation of the net weight ends normally.

  The k-means method is a typical technique for non-hierarchical cluster analysis (clustering). In the k-means method, a prototype that is representative of each of a fixed number of clusters is given, and clustering is performed by assigning each individual to the closest prototype. Once an individual is assigned, next, a new prototype is calculated from the assigned individual. As described above, by repeating the calculation of the prototype and the assignment of the individual until convergence, an appropriate prototype estimation and data division are performed.

  The algorithm of the k-means method is as follows: 1) randomly set the centers of k clusters, 2) assign each individual to the nearest center, 3) recalculate the center for each cluster, and all cluster centers If there is no change, the process ends. Otherwise, the process returns to 2).

  Next, with reference to FIG. 8, the details of the calculation process performed in each of the above-described processes will be described. FIG. 8 is an explanatory diagram for explaining variables used in the calculation process performed in each of the above-described processes. Details of the variables shown in each row will be described in order from the top.

_Tc is a variable indicating the current time. y (t) is an actual measurement value of the objective variable at time t. y (t) is a quality evaluation index for the final process in the manufacturing process. x _i (t) is an actual measurement value at the time t of the explanatory variable. x _i (t) is an observed value related to the i-th characteristic of the intermediate process in the manufacturing process. N is a variable indicating the number of all explanatory variables.

Y _ts (t) is an actual measurement value of the smoothed data at the time t of the objective variable. y _ce (t) is an actual measurement value of the causal variation at time t of the objective variable. y ^ _ts (t) (in this specification, y ^ indicates a symbol obtained by adding a hat symbol ^ on y) is a predicted value of the smoothed data at the time t of the objective variable. y ^ _ce is a predicted value of the causal variation at time t of the objective variable. y ^ (t) is a predicted value of the objective variable at time t. The hat symbol ^ is added to the predicted value.

N _s is the number of explanatory variables that directly affect the objective variable. T _d is a delay time when the target variable series is smoothed. T _l is the number of past measurement values used to predict the next point of the smoothed objective variable. Note that T _l dates back to the past, starting from the value immediately before the current point. T _s is the time when it is assumed that the change in the smoothed objective variable is stable. α _i (t) is a weight set in which the actual measurement value before the past i hours at time t in the smoothing fluctuation prediction affects the current value.

N _h is the number of intermediate layer neurons of the radial basis network (RBFN). r is the radius of the basis function of each neuron in the intermediate layer of RBFN. w _i is a weight set of a network connecting the i-th intermediate layer neuron and the output layer neuron of RBFN. ε is a parameter for reliably executing an operation when calculating a weight set of a network connecting the intermediate layer and the output layer of RBFN. c _ij is a central value corresponding to the j-th objective variable at the central point of the i-th neuron in the intermediate layer of RBFN. z _i (t) is an output value at time t of the i-th neuron in the intermediate layer of RBFN.

Of the above variables, T _d , T _s , N _h , r, w _i , ε, and C _ij are designated as parameters.

  First, a calculation formula for performing the characteristic separation process in the characteristic separation function will be described. Expression 2 is an expression for smoothing the measured objective variable and calculating smoothed data.

  The residual between the smoothed objective variable extracted by Equation 2 and the measured objective variable is extracted using Equation 3 to calculate the causal variation.

  The characteristic separation process is performed by applying the above Equations 2 and 3 to the measured objective variable.

Next, the prediction learning process by each prediction model is demonstrated. In modeling by smoothed fluctuation prediction of smoothed data (trend fluctuation), autoregressive analysis is used to formulate the prediction relationship of data of past T _s time starting from time T _c −1 one unit time before. To do. At this time, the relational expression for predicting the time t is formulated as Expression 4 using the regression coefficient α _i (T _c ).

This is created for the past T _s unit time to make simultaneous equations. The simultaneous equations thus obtained are solved to calculate the regression coefficient α _i (T _c ) for the interval of the past T _s unit time.

  Next, predictive learning processing based on difference data will be described. In the modeling based on the difference fluctuation prediction of the difference data, in step 1, an effective explanatory variable group that directly affects the objective variable is selected from the observed explanatory variable group by using causal structure acquisition.

  After that, in step 2, a radial basis function network is constructed which takes valid explanatory variable groups as inputs and outputs variations due to causality.

Specifically, in step 2-1, for valid explanation variable group selected in step 1, the variation due to the measured value and the causal past T _s times a set of RBFN the input-output relationship. Thereafter, in step 2-2, representative points that match the number of intermediate layer neurons are selected from the set created in step 2-1 by the k-means method.

  In addition, instead of selecting representative points by the k-means method, using the target value of the objective variable or the error in the input / output relationship of the RBFN, the points to improve the prediction accuracy as shown in FIG. 12 are extracted and learned. May be performed. This method will be described later in another embodiment.

  Thereafter, as step 2-3, the representative point selected in step 2-2 is positioned as the center point of each neuron in the intermediate layer, and according to the RBF having a predetermined radius r, the data set of all explanatory variable groups is expressed by equation 5 Is used to calculate the output value of the intermediate layer.

Then, finally, as step 2-4, the net weight w _i between the intermediate layer and the output layer is calculated thereafter. At this time, the normalization parameter ε is used so that the calculation of the net weight ends normally.

  Next, processing of a prediction function in the prediction integration unit 76 that predicts smoothed fluctuations of smoothed data using a smoothed fluctuation model will be described. Expression 6 is an expression for predicting smoothed data at the next point.

  Next, the objective variable group determined to be effective in the causal structure acquisition is input to RBFN, and the difference data of the next point is predicted by Expression 7.

  Finally, the prediction results based on the two learning models obtained by Expression 6 and Expression 7 are added together using Expression 8.

  As a result, a more accurate prediction result can be derived.

  Next, a procedure when the cause estimation unit 78 performs cause estimation processing will be described. First, as step 1, smoothing data at the next point is calculated by applying smoothing fluctuation prediction according to equation (9).

  Next, as step 2, the difference data at the past N points is calculated and applied to Equation 10 to calculate the standard deviation σ.

  Thereafter, as step 3, using the center of fluctuation and the standard deviation obtained in step 1, for example, a temporal abnormality cause estimation reference value (LTAL, HTAL) having a width of 3σ is defined by equations 11 and 12.

  Further, by using the standard deviation of the difference data with the fluctuation obtained in Step 1 as the center, for example, the sudden abnormality cause estimation reference value (LAAL, HAAL) having a width of 3σ is defined by Expression 13 and Expression 14. To do.

  Finally, as Step 4, if the measured current fluctuation exceeds the standard for the two types of abnormal cause estimation standard values obtained in Step 3, the current product is NG and not exceeded In this case, it is determined that the current product is OK.

  By applying the above calculation, processing performed in each functional block is executed.

  Next, an example of a result of performing a prediction process after performing each function by performing the above calculation process and performing a learning process will be described with reference to FIGS.

  14 to 17, the learning and prediction processes described above are performed by smoothing: 45, prediction order: 15, learning set: 100, RBFN intermediate layer neuron number: 10, RBFNσ: 0.3, and RBFNλ: 0.1. It is a graph which shows the result applied with a parameter. 14 to 17, the measured value 105, the smoothed measured value 106, the smoothed predicted value 107, the complemented predicted value 108, the normal / abnormal flag 109, and the sudden abnormality cause estimation reference value 110 (the sudden abnormality cause estimation). Among the reference lower limit LAAL110a and the sudden abnormality cause estimation reference upper limit: HAAL110b), some are selected and shown.

  In FIG. 14, the measured value 105, the smoothed measured value 106, and the sudden abnormality cause estimation reference values 110a and 110b are shown in the graph. Since the actual measurement value 105 includes many variations, it can be seen that a numerical value exceeding the control limit 110a is detected 20 times or more.

  In FIG. 15, the smoothed actual measurement value 106, the smoothed predicted value 107, and the sudden abnormality cause estimation reference values 110a and 100b are shown in the graph. It can be seen that a numerical value exceeding the sudden abnormality cause estimation reference value 110a is detected six times in the complemented predicted value.

  In FIG. 16, the smoothed actual measurement value 106, the complemented predicted value 108, and the sudden abnormality cause estimation reference values 110a and 100b are shown in the graph. It can be seen that in the supplemented predicted value, a numerical value exceeding the sudden abnormality cause estimation reference value 110a is detected 12 times.

  In FIG. 17, the smoothed actual measurement value 106, the normal / abnormal flag 109, and the sudden abnormality cause estimation reference values 110a and 100b are shown in the graph. It can be seen from the data of FIGS. 14 to 16 that the abnormality flag has been detected nine times.

  As described above, the prediction device 2 according to the present embodiment is a characteristic value prediction device 2 that predicts a characteristic value at a certain point of time from a plurality of sensor information acquired in the past, and the plurality of sensor information acquired in the past. An explanatory variable acquisition unit 14 that acquires as an explanatory variable, an objective variable acquisition unit 12 that acquires characteristic values to be predicted from multiple pieces of sensor information as target variables along a time series, and a plurality of objective variables along a time series Is smoothed within a predetermined range, and a smoothed data extraction unit 32 that extracts a target variable smoothed from variations as a trend variation, and removes the target variable smoothed from a plurality of target variables in time series. Thus, the difference data extraction unit 34 that extracts the cause-specific variation determined to have a correlation with the characteristic value, and the difference data that creates the smoothing variation model 54 from the cause-specific variation. A modeling unit 44, and a difference fluctuation prediction unit 72 for predicting a characteristic value at some point from the difference variation model 54.

  According to the above configuration, by extracting the smoothed objective variable from the objective variables based on a plurality of sensor information acquired in the past, it is possible to extract the cause specifying variation determined to have a correlation with the characteristic value. Therefore, it is possible to generate a prediction model from which the cause-specific variation is removed and a prediction model based on the cause variation, create a model by applying an optimum method, and predict characteristic values.

  Further, in the prediction device 2 of the present embodiment, a radial basis function network having an intermediate layer neuron is used as the difference data modeling unit 44, and the difference data modeling unit 44 uses the fluctuation of the variation factor characteristic value and the difference data. Learning data generating means for generating learning data including fluctuations, representative point selecting means for selecting a representative point matching the intermediate layer neuron from the learning data by the k-means method, and an intermediate layer based on the selected center point Network construction means for constructing the radial basis function network by calculating the weight value of the neuron based on the calculated output value of the intermediate layer.

  According to the above configuration, since the intermediate layer neuron as a representative point is selected by the k-means method, the learning of the radial basis function network can be performed in a short time.

  In the prediction device 2 of the present embodiment, the characteristic value at a certain point in time is a quality characteristic in the final process of the production line, and the plurality of pieces of sensor information acquired in the past are sensor information by the inspection unit of the production line. The final process is a process that cannot be corrected after the process in the production line is completed, and the explanatory variables are sensors acquired by the inspection means after the completion of each process of the production line by the sensor devices. Preferably, the objective variable is sensor information acquired by the inspection means after completion of the final process.

  According to the above configuration, the characteristic value of the final process that cannot be corrected after completion can be predicted based on the sensor information up to the end stage of the production line that can be redone. The quality can be determined in advance and corrected.

  Moreover, in the prediction apparatus 2 of this embodiment, based on the smoothing prediction part 74 which performs prediction from the smoothing fluctuation model 54, the prediction result by the smoothing prediction part 74, and the prediction result by the difference fluctuation prediction part 72, It is preferable to include a prediction integration unit 76 that predicts the characteristic value of the final process.

  According to the above configuration, the prediction results based on the two prediction models are added together based on the smoothed variation model 54 and the difference variation model 52 that have been generated by the optimal modeling method, thereby making it more accurate. Predictions can be made.

  In addition, the prediction device 2 of the present embodiment preferably further includes a parameter setting unit 60 that dynamically sets a threshold value based on the prediction result by the smoothing prediction unit 74 and the prediction result by the difference fluctuation prediction unit 72.

  According to the above configuration, the judgment criteria for abnormality are set for each of the change in quality characteristics over time and the abnormality in quality characteristics due to sudden change with respect to the control limit set in the straight line. Can do.

  Further, in the prediction device 2 of the present embodiment, a description highly related to the characteristic value of the final process from the characteristic value of the final process predicted by the prediction integration unit 76 and at least one of the difference variation model 52 and the explanatory variable. It is preferable to determine the variable.

  According to the above configuration, it is possible to identify the cause of the defect by identifying the cause that has affected the characteristic value of the final process, and to improve the process, environment, settings, and the like. In the above description, regarding the application example to the board mounting quality characteristic prediction, the objective variable is the fillet length of the component A, the explanatory variables are the solder amount of the component A, the vertical deviation of the component A, and the lateral deviation of the component A. Although the case has been described, the present invention is not limited to this. Applying to power consumption prediction, the objective variable may be the basic electric energy, and the explanatory variable may be the temperature at measurement point A, the temperature at measurement point B, and the temperature at measurement point C. If an objective variable to be predicted and an explanatory variable that is considered to affect the prediction are appropriately set and a prediction model that takes into account the causal relationship can be formed, the same effects as in the present embodiment can be obtained.

  However, when applied to the prediction of the objective variable of the final process in a production process such as board mounting as in this embodiment, the state after the reflow process is predicted before executing the process in the final process, here the reflow process. Therefore, it is possible to prevent the occurrence of defective products by determining the quality of the substrate before the reflow process and correcting the position of the component before the reflow process, which is particularly effective.

[Embodiment 2]
The prediction system of the present embodiment has a configuration in which part of the modeling processing method in the non-linear prediction of the difference fluctuation is different from the system shown in FIGS. 1 to 17, and the other configurations are the same. . Specifically, among the parameters shown in FIG. 8, was not used in the above embodiment, the number N _k, and (2) the target value T _a of RBFN intermediate layer neurons located in the vicinity of (1) a target value newly It is configured to be used as a simple parameter.

  In addition, the same code | symbol is attached | subjected to the structure which has the function similar to the structure demonstrated in the said embodiment, and the description is abbreviate | omitted.

  In the above configuration, the predictive learning process based on the difference data will be described as follows based on FIG. 12 and FIG.

  The procedure of differential data modeling processing in this embodiment is as follows.

  In the modeling based on the difference fluctuation prediction of the difference data, in step 1, an effective explanatory variable group that directly affects the objective variable is selected from the observed explanatory variable group by using causal structure acquisition.

  After that, in step 2, a radial basis function network is constructed which takes valid explanatory variable groups as inputs and outputs variations due to causality.

Specifically, in step 2-1, for valid explanation variable group selected in step 1, the variation due to the measured value and the causal past T _s times a set of RBFN the input-output relationship. Thereafter, in step 2-2, N _k upper center points whose objective variables are close to the target value are selected from the set of input / output relationships created in step 2-1.

  Thereafter, as step 2-3, the representative point selected in step 2-2 is positioned as the center point of each neuron in the intermediate layer, and according to the RBF having a predetermined radius r, the data set of all explanatory variable groups is expressed by equation 5 Is used to calculate the output value of the intermediate layer.

Thereafter, as step 2-4, the net weight w _i between the intermediate layer and the output layer is calculated thereafter. At this time, the normalization parameter ε is used so that the calculation of the net weight ends normally.

Thereafter, as Step 2-5, an error between the RBFN constructed in Step 2-4 and the input / output relationship of the RBFN created in Step 2-1 is calculated. Further, as step 2-6, N _h -N _k center points with large errors calculated in step 2-5 are selected.

Finally, as Step 2-7, the center point selected in Step 2-6 is positioned as a new center point of RBF added to the RBFN constructed in Step 2-4, and is the same as Step 2-3 and Step 2-4. The radius r is determined, the output value of the intermediate layer is calculated, and the process of calculating the weight w _i is repeated.

  With the above processing, as shown in FIG. 12, by extracting an arbitrary point for which the prediction accuracy is to be improved, a point for which the prediction accuracy is to be improved can be extracted and the prediction accuracy of a desired point can be improved. Therefore, it is possible to accurately determine whether or not to implement a measure by accurately predicting a point to be finally determined.

  As described above, the prediction device 2 of the present embodiment uses a radial basis function network having an intermediate layer neuron as the difference data modeling unit 44, and the difference data modeling unit 44 uses the fluctuation of the variation factor characteristic value and the above-described variation factor characteristic value. Learning data generating means for generating learning data including fluctuations of difference data, first center point selecting means for selecting a center point of an intermediate layer neuron from the learning data based on a target value of the difference data, and a selected center Network construction means for constructing the radial basis function network by calculating an output value of the intermediate layer based on the points, and calculating a weight value of the neuron based on the calculated output value of the intermediate layer, and the network construction By applying the above learning data to a radial basis function network constructed by means of And a second center point selecting means for selecting a new center point of the intermediate layer neuron based on an error between the calculated and calculated estimated value and the fluctuation of the difference data. The output value of the intermediate layer is calculated based on the center point selected by the second center point selection means, and the weight value of the neuron is calculated based on the calculated output value of the intermediate layer.

  According to the above configuration, it is possible to extract a point to be improved in prediction accuracy and improve the prediction accuracy of a desired point, so that a measure is implemented by accurately predicting a point to be finally determined. Alternatively, it is possible to accurately determine that the operation is not performed.

  For example, if the final quality is predicted in the production line, measures are taken by accurately predicting whether the final quality is good or not by accurately predicting values near the control limit value. Since it is possible to accurately determine whether or not to implement, the cost of measures for the product can be optimized.

  As another specific example, in the office or factory, the future power consumption is predicted, and if there is a possibility that the demand contracted with the power supplier may be exceeded, the operation of the equipment that consumes power is temporarily suspended. In the case of a system that takes measures such as stopping automatically, it is possible to accurately determine whether measures are to be implemented or not by accurately predicting values near the preset demand power, so that demand exceeds It is possible to minimize the impact on the business.

  The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope shown in the claims, and embodiments obtained by appropriately combining technical means disclosed in different embodiments. Is also included in the technical scope of the present invention.

  Finally, each block of the prediction system 1 of the present invention, in particular, the learning unit 20 and the prediction unit 70 may be configured by hardware logic, or may be realized by software using a CPU as follows.

  That is, the prediction system 1 includes a CPU (central processing unit) that executes instructions of a control program that realizes each function, a ROM (read only memory) that stores the program, a RAM (random access memory) that expands the program, A storage device (recording medium) such as a memory for storing the program and various data is provided. An object of the present invention is to provide a recording medium in which a program code (execution format program, intermediate code program, source program) of a control program of the prediction system 1 that is software that realizes the above-described functions is recorded so as to be readable by a computer. This can also be achieved by supplying the prediction system 1 and reading and executing the program code recorded on the recording medium by the computer (or CPU or MPU).

  Examples of the recording medium include tapes such as magnetic tapes and cassette tapes, magnetic disks such as floppy (registered trademark) disks / hard disks, and disks including optical disks such as CD-ROM / MO / MD / DVD / CD-R. Card system such as IC card, IC card (including memory card) / optical card, or semiconductor memory system such as mask ROM / EPROM / EEPROM / flash ROM.

  Moreover, the prediction system 1 may be configured to be connectable to a communication network, and the program code may be supplied via the communication network. The communication network is not particularly limited. For example, the Internet, intranet, extranet, LAN, ISDN, VAN, CATV communication network, virtual private network, telephone line network, mobile communication network, satellite communication. A net or the like is available. Also, the transmission medium constituting the communication network is not particularly limited. For example, even in the case of wired such as IEEE 1394, USB, power line carrier, cable TV line, telephone line, ADSL line, etc., infrared rays such as IrDA and remote control, Bluetooth ( (Registered trademark), 802.11 wireless, HDR, mobile phone network, satellite line, terrestrial digital network, and the like can also be used. The present invention can also be realized in the form of a computer data signal embedded in a carrier wave in which the program code is embodied by electronic transmission.

  As described above, the prediction system 1 and the prediction device 2 of the present invention have a correlation with the characteristic value by removing the smoothed objective variable from the objective variable based on a plurality of sensor information acquired in the past. Since the identified cause-specific variation can be extracted, a prediction model that eliminates the cause-specific variation and a prediction model based on the cause variation are generated, and the model is created by applying the optimal method to predict the characteristic value. Since it can be performed, it can be suitably used as a predicting device for final quality characteristics in a production line or a predicting device for predicting various other events based on sensor information.

1, showing an embodiment of the present invention, is a block diagram illustrating a main configuration of a prediction system. FIG. It is the schematic shown about the flow of the process at the time of performing a learning process and a prediction / determination process in said prediction system. In the above prediction system, it is a flowchart showing the flow of processing when executing the characteristic separation processing and the prediction learning processing. It is a flowchart which shows the flow of a process at the time of performing a parameter adjustment process in said prediction system. It is a flowchart which shows the flow of a process at the time of performing a prediction process in said prediction system. In said prediction system, it is a flowchart which shows the flow of a process at the time of performing a cause estimation process. In said prediction system, it is a flowchart which shows an example of a difference data modeling process. In said prediction system, it is a figure which shows the table | surface explaining the list | wrist of the variable used for the various calculations used when processing. It is a figure which shows the structure of the basis function network prediction model used for a difference data modeling part in said prediction system. It is a figure showing the Gaussian function used in said difference data modeling part. FIG. 11A and FIG. 11B are diagrams showing an outline of the arrangement of default functions in the basis function network prediction model. It is a figure explaining selection of the center point in said basis function network prediction model. It is a figure which shows the graph explaining the relationship between an objective variable and explanatory variables in said prediction system. It is a figure which shows the graph of the result predicted using said prediction system, and the upper limit and lower limit of a measured value, a smoothing measured value, and a sudden abnormality cause estimation standard are shown. It is a figure which shows the graph of the result predicted using said prediction system, and shows the smoothing actual value, smoothing prediction value, and the upper limit and minimum of sudden abnormality cause estimation reference | standard. It is a figure which shows the graph of the result predicted using said prediction system, and shows the smoothing actual value, the complemented prediction value, and the upper limit and the minimum of a sudden abnormality cause estimation reference | standard. It is a figure which shows the graph of the result predicted using said prediction system, and the upper limit and lower limit of a measured value, a normal / abnormal flag, and a sudden abnormality cause estimation reference | standard are shown. In said prediction system, it is a flowchart which shows another example of a difference data modeling process.

Explanation of symbols

1 Time Series Data Analysis System 2 Time Series Data Analysis Device (Time Series Data Analysis Device)
10 Acquisition Unit 12 Objective Variable Acquisition Unit (Objective Variable Acquisition Means)
14 Explanation variable acquisition part (Explanation variable acquisition means)
20 learning unit 30 characteristic separating unit 32 smoothed data extracting unit (smoothed data extracting means)
34 Difference data extraction unit (difference data extraction means)
40 Prediction learning unit 42 Smoothing data modeling unit (smoothing data modeling means)
44 Differential data modeling unit (differential data modeling means)
52 Differential Fluctuation Model (Differential Data Fluctuation Model)
54 Smoothing fluctuation model (smoothing data fluctuation model)
60 Parameter setting unit (threshold setting means)
70 Prediction Unit 72 Difference Variation Prediction Unit (Difference Variation Prediction Unit)
74 Smoothing fluctuation prediction unit (smoothing fluctuation prediction means)
76 Predictive Integration Unit (Predictive Integration Unit)
78 Cause estimation part (variation factor estimation means)
80 process processing equipment 82 printing inspection equipment (explanatory variable inspection equipment)
84 Deviation inspection device (Explanatory variable inspection device)
86 Fillet length inspection device (objective variable inspection device)
90 Process management device 92 Printing device 94 Mounting device 96 Reflow device

Claims (13)

A time-series data analysis device that analyzes time-series data of prediction target characteristic values in the system based on detection result information of a plurality of types of characteristic values in a predetermined system,
Time-series data acquisition means for acquiring time-series data consisting of past detection results of the prediction target characteristic value;
Smoothed data acquisition means for calculating smoothed data obtained by smoothing the time series data according to a predetermined standard;
Difference data acquisition means for calculating a difference between the time series data and the smoothed data as difference data;
Smoothing data modeling means for identifying a correspondence relationship between the passage of time and the fluctuation of the smoothed data as a smoothed data fluctuation model by analyzing past fluctuations of the smoothed data;
Difference data that specifies the type of the characteristic value that affects the fluctuation of the difference data as a fluctuation factor characteristic value, and specifies the correspondence between the fluctuation of the fluctuation factor characteristic value and the fluctuation of the difference data as a difference data fluctuation model A time-series data analysis apparatus comprising a modeling means. As the differential data modeling means, a radial basis function network with intermediate layer neurons is used,
The differential data modeling means is
Learning data generating means for generating learning data including fluctuation of the fluctuation factor characteristic value and fluctuation of the difference data;
First center point selection means for selecting a center point of the intermediate layer neuron based on the target value of the difference data from the learning data;
A network construction means for constructing the radial basis function network by calculating an output value of the intermediate layer based on the selected center point, and calculating a weight value of the neuron based on the calculated output value of the intermediate layer;
By applying the learning data to the radial basis function network constructed by the network construction means, an estimated value of the difference data is calculated, and an intermediate layer neuron is calculated based on an error between the calculated estimated value and the fluctuation of the difference data. And a second center point selecting means for selecting a new center point of
The network construction means calculates an intermediate layer output value based on the center point selected by the first and second center point selection means, and calculates a neuron weight value based on the calculated intermediate layer output value. The time-series data analysis apparatus according to claim 1. As the differential data modeling means, a radial basis function network with intermediate layer neurons is used,
The differential data modeling means is
Learning data generating means for generating learning data including fluctuation of the fluctuation factor characteristic value and fluctuation of the difference data;
a representative point selecting means for selecting a representative point matching the intermediate layer neuron from the learning data by the k-means method;
Network construction means for constructing the radial basis function network by calculating an output value of the intermediate layer based on the selected center point, and calculating a weight value of the neuron based on the calculated output value of the intermediate layer,
The time-series data analysis device according to claim 1, comprising:   The variation factor estimating means for estimating the variation factor of the characteristic value at the prediction target time point from the smoothed data variation model and the difference data variation model is further provided. Time series data analysis device.   The time-series data according to any one of claims 1 to 3, further comprising a predicting unit that predicts a characteristic value at a prediction target time point from the smoothed data variation model or the difference data variation model. Analysis device. The prediction means is
A difference data prediction means for obtaining a characteristic value at the time of the prediction target and calculating a predicted value of the difference data based on the difference data fluctuation model;
Smoothed data prediction means for calculating a predicted value of the smoothed data at the prediction target time based on the smoothed data variation model;
The prediction integration means for calculating the prediction value of the characteristic value at the measurement target time point based on the prediction result by the difference data prediction means and the prediction result by the smoothed data prediction means. 5. The time-series data analysis device according to 5. The plurality of types of characteristic values in the predetermined system are sensor information detected in each process of the production line in the production system,
The time series data analysis apparatus according to any one of claims 1 to 6, wherein the time series data to be predicted is time series data of quality characteristics of a production object in the production line.   The apparatus further comprises threshold setting means for setting an upper limit value and a lower limit value of control limit values, which are threshold values for managing quality characteristics of the production object, based on a prediction result by the smoothed data prediction means. Item 8. The time-series data analysis device according to Item 7.   Threshold setting means for setting an upper limit value and a lower limit value of a sudden abnormality cause estimation criterion, which is a threshold value for managing quality abnormality due to sudden change of the production object, from the prediction result by the difference data prediction means. The time-series data analysis apparatus according to claim 7. A time series data analysis device according to any one of claims 1 to 9,
An objective variable inspection device for acquiring sensor information from a sensor and inputting it as time-series data to the time-series data acquisition means,
Explanatory variable inspection device for acquiring sensor information from a sensor and inputting it as a learning parameter to the smoothed data modeling means and the differential data modeling means;
A model creation system comprising: a management device that manages acquisition of the sensor information. A time-series data analysis method for analyzing time-series data of prediction target characteristic values in the system based on detection result information of a plurality of types of characteristic values in a predetermined system,
Obtaining time-series data consisting of past detection results of the prediction target characteristic value;
Calculating smoothed data obtained by smoothing the time series data according to a predetermined standard;
Calculating a difference between the time series data and the smoothed data as difference data;
Identifying the correspondence between the passage of time and the fluctuation of the smoothed data as a smoothed data fluctuation model by analyzing the past fluctuation of the smoothed data;
Identifying the type of the characteristic value that affects the variation of the difference data as a variation factor characteristic value, and identifying the correspondence between the variation of the variation factor characteristic value and the variation of the difference data as a difference data variation model A time-series data analysis method characterized by comprising:   A program for operating the time-series data analysis device according to claim 1, wherein the computer functions as each of the above-described means.   A computer-readable recording medium on which the program according to claim 12 is recorded.

Images