TWI667660B - Intelligent pre-diagnosis and health management system modeling method and computer program product - Google Patents

Abstract

The invention provides a smart pre-diagnosis and health management system modeling method, which comprises a new tree establishing step, a dual-branch modeling step, and a model self-adaptation optimization step, which can be modeled by a dual branch with the increase of monitoring data. The proposed prediction hypothesis model selects the best solution as the benchmark for the optimal decision, which is used for the next forecast and makes the system's prediction result meet the expected target value. The invention also provides a computer program product of a smart pre-diagnosis and health management system, which performs the above-mentioned intelligent pre-diagnosis and health management system modeling method when executing the computer program product.

Description

Intelligent pre-diagnosis and health management system modeling method and computer program product

The invention relates to a smart prognostics and health management (SPHM) system modeling method and a computer program product thereof, in particular to a smart pre-diagnosis system modeling method for optimizing self-adaptation and Its computer program product.

In order to ensure the stability of the production process of the production machine and increase the rate of production, the manufacturing industry must strictly monitor the operation status of the production machine. Using machine pre-diagnosis and health management technology, you can monitor and evaluate the health status of equipment and parts by analyzing machine data, and determine the optimal maintenance or replacement timing based on status to reduce unplanned downtime and reduce maintenance. frequency.

In order to meet quality requirements, the prior art has strict monitoring and observation of key process parameters. The so-called "key process parameters" refer to the factors most relevant to equipment failures. In practice, these factors are monitored as an important indicator for equipment maintenance pre-diagnosis. In order to improve the accuracy of the pre-diagnosis, a number of public technologies have been proposed for various improvements, including the applicant's selection of leading auxiliary parameters in US Patent Application No. US16/001,520, and the maintenance of equipment in combination with key parameters and leading auxiliary parameters. The method of diagnosis is to filter the data collected by the sensor and distinguish it into a set of critical parameters (CP) and other feature parameter sets, and then identify the earliest time in the feature parameter set to influence the key parameters in advance. Leading associated parameters (LAP), and further use the key parameter (CP) set and the leading auxiliary parameter (LAP) to establish a device maintenance pre-diagnosis model that effectively improves the early warning capability.

However, when performing equipment maintenance pre-diagnosis model modeling, the prior art technology must accumulate a certain amount of monitoring data and maintenance records to begin modeling. However, in the early days when new machine equipment was introduced into the pre-diagnostic system, engineers often could not quickly and effectively find important parameters and establish predictive models from a small number of monitoring data without maintenance records. Therefore, there is a need to provide a smart pre-diagnosis and health management system modeling method and computer program product thereof to overcome the shortcomings of the above-mentioned prior art.

One of the objects of the present invention is to solve the shortcomings in the initial stage of introducing a new machine equipment into a pre-diagnostic system, and it is difficult to find important parameters and establish a prediction model through a small amount of monitoring data without maintenance records.

Another object of the present invention is to select a best model from a bi-branched model to optimize the decision through the two-branch modeling method, so that the pre-diagnosis is consistent with the prediction result of the health management system. Expected target value.

In order to achieve the above object, the present invention provides a smart pre-diagnosis and health management system modeling method by a smart pre-diagnosis and health management system with a plurality of reference hypothesis models built therein. The method includes:

(S1) a new tree establishing step, wherein at least one analysis tree node (Object) is defined according to a to-be-monitored machine, and each analysis tree node (Object) obtains a monitoring data via at least one monitoring point;

(S2) A two-branch modeling step is performed by performing a data pre-processing step a on a branch 1 to convert the monitoring data into a specified feature format and performing a similarity analysis to select the highest similarity from the reference hypothesis models. And a reference hypothesis model exceeding a specified threshold is used as a branch 1 prediction hypothesis model of the analysis tree node; and a data pre-processing step b is performed on the monitoring data in a branch 2 to confirm the analysis tree node by using a causality check ( Object) a key parameter (CP) and a plurality of related parameters (AP) and construct a hypothesis model applicable to the machine to be monitored as a branch 2 prediction hypothesis model of the analysis tree node;

(S3) The model adaptive optimization step, if the monitoring data is continuously generated, after the judgment of the analysis tree node (Object) is "whether the machine can not continue the maintenance", if the judgment result is "Yes" And selecting an optimal solution from the branch 1 prediction hypothesis model or the branch 2 prediction hypothesis model constructed by the dual branch modeling step as a benchmark for the optimization decision of the analysis tree node (Object), the benchmark Used for the next forecast and the system's forecast results are in line with the expected target value.

According to an embodiment of the present invention, in the new tree establishing step, the machine to be monitored may be a new machine without a maintenance record.

According to an embodiment of the present invention, in branch 1 of the dual branch modeling step, the specified feature format may have the same feature format as before the reference hypothesis model is modeled.

According to an embodiment of the present invention, after the model optimization step is specified, the monitoring data may be further converted into the specified feature format and updated to the smart pre-diagnosis and health management system. Some reference hypothesis models.

According to an embodiment of the present invention, in the step of optimizing the model, the method further includes: acquiring at least one maintenance record, wherein the set of maintenance records includes at least one maintenance status value of the monitoring point, so that the monitoring data and the maintenance record are one The relationship to one.

According to an embodiment of the present invention, in the branch 1 of the two-branch modeling step, the similarity analysis may adopt at least one selected from the Eucliled Distance, the Mahalanobis Distance, and the Mahalanobis Distance. Manhattan Distance, Minkowski distance, Cosine Similarity, and combinations thereof are used as a quantification method for this similarity analysis.

According to an embodiment of the present invention, in the branch 1 of the two-branch modeling step, when the result of the similarity analysis does not exceed the specified threshold, the key parameter (CP) confirmed by the branch 2 may be first used. And the feature matrix composed of the relevant parameters (AP) is used as the basis for modeling the branch 1 prediction hypothesis model, and after obtaining a maintenance record label in the model adaptive optimization step, the branch is reconstructed according to a supervised learning method. Forecast hypothesis model.

According to an embodiment of the invention, the supervised learning method may include at least one selected from the group consisting of a support vector machine (SVM), a regression analysis (Regression), a random forest (Random Forest), and combinations thereof. .

According to an embodiment of the present invention, in the model adaptive optimization step, when the branch 1 prediction hypothesis model is superior to the branch 2 prediction hypothesis model for m consecutive times, the branch 1 prediction hypothesis model may be set as the optimal solution; When the branch 1 prediction hypothesis model is not better than the branch 2 prediction hypothesis model for m consecutive times, the branch 2 prediction hypothesis model can be specified as the optimal solution.

According to an embodiment of the present invention, in the model adaptive optimization step, the m value may be a positive integer greater than 2.

The invention further provides a computer program product of a smart pre-diagnosis and health management system, which can complete the above-mentioned intelligent pre-diagnosis and health management system modeling method when executing the computer program product.

Therefore, the present invention can start analysis from a small amount of analysis tree node monitoring data without maintenance records, find important parameters and use the analysis tree node bifurcation modeling to establish a prediction model. As the analysis tree node monitoring data and the corresponding maintenance records continue to increase, through the model adaptive optimization step, the best model is selected from the analysis tree node double-branch model to select the best solution (golden model), analyze and know the important Monitor the impact of the parameters on the operating state of the machine, and then find out the best time to repair and replace the parts to ensure the stability of the production process and increase the productivity and utilization rate.

The detailed description and technical content of the present invention will now be described as follows:

1A is a structural diagram of a smart pre-diagnosis system according to an embodiment of the present invention. The smart pre-diagnosis and health management (SPHM) system 10 includes at least an analytic engine service manager (AESM) module 20 A smart prediction and health management object analysis tree (SPHM-OAT) module 30, a machine learning library module 40, and a file system module 50.

In order to make the application of the system of the present invention more scalable, the intelligent pre-diagnosis and health management system 10 may further include an expansion module that is coupled to the intelligent prediction and health management object analysis tree module 30. The expansion module can include a first exchangeable application interface 60a, a second exchangeable application interface 60b, and an exchangeable driver interface 60c. The first exchangeable application interface 60a is used to connect a An external machine learning module 70, the second exchangeable application interface 60b is connected to an external reference model module 80, and the exchangeable driver interface 60c is connected to an external data collection drive (EDCD) 90. The original data set in the database 91 of the equipment to be monitored is obtained.

The analysis engine service management module 20 is the core of the intelligent pre-diagnosis and health management system 10 of the present invention, and can control the state of each component in the intelligent prediction and health management object analysis tree (SPHM-OAT) module 30.

The intelligent predictive and health management object analysis tree module 30 is coupled to the analysis engine service management module 20 and includes a plurality of analysis trees (OATs), each of which includes a plurality of analysis tree nodes (Object), each An analysis tree node corresponds to a key parameter (CP) and a plurality of related parameters (AP), and the SPHM Health Indicator (SPHM-HI) timely reflects the health of each node of the analysis tree. State, make early warning and health management. The Health Indicator (SPHM-HI) is scalable, and examples of basic items may include Next N-Run Fail (NRF) indicators, Remaining Useful Life (RUL) indicators for equipment critical components. General Health Indicators (HI), and other similarly related health indicators, are not described here because the functions, types, actual quantification and analysis methods of these health indicators are well known to those familiar with the art.

The data sources of the key parameters (CP) and related parameters (AP) may be information obtained by the sensor, or may be aggregated by key parameters (CP) and other related parameters (AP) of the analysis tree nodes (Object). Made. Each analysis tree node is connected by an object control table (OCB), which is used to store the corresponding operation result of the analysis tree node in the analysis process, and has the effects of regular backup and restoration. In this way, if a disaster event occurs during the analysis process, the object control table can quickly perform the reply operation, obtain the state of the analysis tree node from the last checkpoint, and then recursively, from the sibling node. The (sibling node) goes to the parent node and continues the hierarchical integration operation analysis from bottom to top until the analysis of the highest level analysis tree node (ie, root) is completed.

Accordingly, the intelligent pre-diagnosis and health management system 10 of the present invention can timely reflect through the analysis tree nodes under the premise that the monitoring data source is correct and the key parameters (CP) and related parameters (AP) are also selected correctly. The health status of the analysis tree nodes is well prepared for early warning and health management.

The Intelligent Prediction and Health Management Object Analysis Tree (SPHM-OAT) module 30 is responsible for workflow management on the analysis tree nodes in addition to managing the analysis trees representing the corresponding complex class machines. The so-called "workflow" is managed by a mapping table, and may include a stacked data preprocessing layer, a data hypothesis layer, and a data ensemble layer. The level, order and actual work content of the workflow can be adjusted according to needs, and is not limited to the above.

The mapping table is operated by a table driven mechanism, and at least the selected one of the machine learning library modules 40 connected to the intelligent prediction and health management object analysis tree module 30 is selected from the table according to a preset working method. A suitable algorithm is used for the above-mentioned workflows such as the data pre-processing layer, the data hypothesis layer, or the data learning layer. For example, the algorithm applicable to the data pre-processing layer may include an algorithm with feature selection ability such as feature selection algorithm or feature extraction algorithm; the algorithm applicable to the data hypothesis layer may be Including regression algorithm, autoregressive integrated moving average model (ARIMA) algorithm, relative strength index (RSI) algorithm or other algorithms with predictive ability; The working method of the layer is to vote by constructing a set of multiple hypothesis models specified by the mapping table, or according to the hierarchical integrated operation specified by the current analysis tree. In addition, the analysis engine service management module 20 also controls the workflow of each analysis tree according to the mechanism of the mapping table.

The file system module 50 can be used as a place for the system to write back files and/or save files. The above-mentioned "files" can include, for example, the life cycle of the analysis trees in the intelligent prediction and health management object analysis tree module 30. Quantitative analysis information, or a preset reference hypothesis model set, a set of feature sample data before modeling, or a backup material when the system fails in the calculation process, or a reference hypothesis to which each analysis tree node belongs, to provide when necessary The intelligent prediction and health management object analysis tree module 30 requires information.

If necessary, the system of the present invention can be expanded by connecting the external device to the external device. For example, when the existing information of the machine learning library module 40 is insufficient, the first exchangeable by the expansion module can be The application interface 60a is connected to the external machine learning module 70 to expand the functionality of the existing machine learning library module 40; or the external reference model can be connected by the second exchangeable application interface 60b of the expansion module. The module 80 is configured to expand the hypothesis model indicator of the mapping table of the intelligent prediction and health management object analysis tree module 30 and participate in the selection and deployment of the external hypothesis model of the manual mode; The switch driver interface 60c is connected to an external data collection drive device 90. The external data collection drive device 90 is connected to the external data library 91. Therefore, the external data collection drive device 90 can obtain the device stored in the device to be monitored. The original data of the external database 91.

Next, the intelligent pre-diagnosis and health management system modeling method of the present invention will be described by way of an example.

Firstly, in the initial stage of introducing the intelligent pre-diagnosis and health management system of the present invention in a new machine, the monitoring data of the analysis tree node having no maintenance record is obtained from scratch, and the monitoring data can be It is raw data, and each monitoring data includes a plurality of monitoring parameters.

After completing the above initialization settings, the dual branch modeling mode of the analysis tree object is turned on. Please refer to "Figure 1B", "Figure 2A" and "Figure 2B". The bifurcation modeling of the analysis tree object is the branch 1 prediction hypothesis model modeling and the branch 2 prediction hypothesis model modeling.

Branch 1 prediction hypothesis model modeling

The flow of the branch 1 modeling of the present invention includes the data pre-processing step a of step 200a and the similarity analysis step of step 300a, and performs the following step 400a according to the above two steps, that is, the modeling of the branch 1 prediction hypothesis model And forecasting.

The data pre-processing step a of step 200a converts the monitoring data of the analysis tree object of the new machine into the same feature format as before the modeling of the built-in reference hypothesis model of the intelligent pre-diagnosis and health management system 10.

The similarity analysis step of step 300a takes the reference hypothesis model index built in the mapping table as a reference, and selects the feature similarity of the model according to the similarity index according to the feature data sample before each reference hypothesis model modeling. The reference hypothesis model of the value is used as the "branch 1 prediction hypothesis model" of the analysis tree object (step 400a).

The similarity analysis of step 300a is to quantize the similarity by calculating the feature distance, and the hypothesis model with the highest similarity to the modeling feature of a reference hypothesis model is selected from the built-in reference hypothesis model as the analysis tree. The branch 1 prediction hypothesis model of the object. In this step, the feature distance calculation methods of various spaces can be used to quantify the similarity, and the technique is well known to those skilled in the art, and therefore will not be described here.

However, if the monitoring data of the new machine is lower than the similarity index value between the converted feature data and the feature data before the reference hypothesis model built in the intelligent pre-diagnosis and health management system 10 When the threshold value is specified, it indicates that the reference hypothesis model built into the smart pre-diagnosis and health management system 10 does not have a hypothesis model applicable to the new machine. Then, the key parameter (CP) obtained by the leading auxiliary parameter analysis step of the step 300b of the branch 2 and the feature matrix composed of the related parameters (AP) will be temporarily used as the branch 1 of the analysis tree node. The hypothesis model (step 400a) is characterized by the modeling steps. After the model adaptive optimization step (step 500) is performed and the maintenance record labels are obtained, the branch 1 prediction hypothesis model of the analysis tree object is reconstructed according to the supervised learning method. Finally, a hypothesis model is added to a smart pre-diagnosis and health management (SPHM) object container. The intelligent pre-diagnosis and health management (SPHM) object container is placed in the intelligent prediction and health management object analysis tree module. In the mapping table of 30, the modeled feature sample data corresponding to the smart pre-diagnosis and health management (SPHM) object container is stored in the file system module 50, and the prediction hypothesis model of the branch 1 is automatically completed. expansion. Regarding the supervised learning methods mentioned above, there are commonly supported support vector machines (SVMs), regression analysis (regression), and random forests. The functions, types, and actual quantification of the above methods are The method of analysis is well known to those skilled in the art and will not be described here.

Branch 2 prediction hypothesis model modeling

The branch 2 prediction hypothesis modeling process of the present invention comprises the following steps: performing data pre-processing step b of step 200b and leading auxiliary parameter analysis step of step 300b, and performing the following step 400b, that is, branch 2 prediction according to the above two steps Hypothesis model modeling and prediction.

Step 200b uses the data cleansing step (refer to step 210b of FIG. 2B) to process the monitoring data of the analysis tree object into a usable state. The above data cleaning steps include, but are not limited to, data alignment, missing value processing, error data correction, and other similar related steps. The functions, types, and actual operation methods are well known to those skilled in the art, and are therefore not included herein. Narration.

Then perform the statistical feature processing step (refer to step 220b of FIG. 2B) and the mixed feature processing step (refer to step 230b of FIG. 2B), according to statistical methods, such as: maximum value, minimum value, average value, Variants, kurtosis, skewness, etc., generate various statistical features; and feature extraction methods such as principal component analysis (PCA), independent components analysis (ICA), etc. Generating various linear transformation features; generating hybrid features based on machine learning, such as neural networks (NNs) algorithm, LSSO, and obtaining a hybrid feature Feature set.

Next, the lead assist parameter analysis step is performed as described in step 300b. The collected feature set is first divided into a key parameter set and a non-key parameter set. According to the causality test, for example, the Granger causality test finds the leading leading parameters (LP) of the key parameters (CP) of the analysis tree of the new machine. ).

Regarding the above-mentioned key parameters (CP) and effective leading auxiliary parameters (LAP), in simple terms, a key parameter (CP) variation is caused by an effective leading auxiliary parameter (LAP), that is, the effective leading auxiliary parameter (LAP) This key parameter (CP) has the ability to lead reactions. Then, the prediction hypothesis model is established by using trend modeling techniques such as regression model and autoregressive integrated moving average model, and then the analysis tree node is constructed by ensemble learning ( Branch 2 prediction hypothesis model of object). Finally, a hypothesis model is added to a smart pre-diagnosis and health management (SPHM) object container. The intelligent pre-diagnosis and health management (SPHM) object container is placed in the intelligent prediction and health management object analysis tree module. In the mapping table of 30, the modeled feature sample data corresponding to the smart pre-diagnosis and health management (SPHM) object container is stored in the file system module 50, and the analysis tree node is completed. The branch 2 prediction hypothesis model is automatically expanded. The above-described methods of feature generation, feature transformation, trend modeling techniques, regression models, ARIMA models, and integrated learning are well known to those skilled in the art and are not described herein.

Dual branch modeling embodiment

Please refer to FIG. 1B for a flowchart of a modeling method of the intelligent pre-diagnosis system according to an embodiment of the present invention. In this embodiment, it is assumed that branch 2 is the best solution in the past training experience, so in step 110 of the first step, the branch 2 prediction model will be preferentially designated as the golden model based on the past training experience.

Then, step 120 is performed to obtain the first n monitoring data sets X of the analysis tree nodes of the new machine device, wherein the monitoring data X of each group of analysis tree nodes includes a plurality of monitoring values. Xij, where i is used to indicate the i-th monitoring parameter, and j is used to indicate the j-th monitoring data. The monitoring data X of the analysis tree nodes corresponds to the monitoring point i in a one-to-one manner.

Next, the data pre-processing steps 200a and 200b of the dual-branch modeling are implemented. The purpose of the data pre-processing steps 200a and 200b of the double branch of the analysis tree node above is to filter invalid or non-influenced monitoring data, and convert the usable monitoring data into branches 1 and/or for branches. 2 The data format of the subsequent steps. For example, the sensor on the particulate filter on a MOCVD machine can convert the monitoring data into a data format suitable for the subsequent similarity analysis step 300a through the data pre-processing step 200a; or through the data pre-processing step 200b converts the monitoring data into a data format for subsequent application of the leading auxiliary parameter analysis step 300b.

The pre-processing step 200b of the pre-processing step 200a and the branch 2 of the branch 1 will be separately described below.

Please refer to "Fig. 2A" for the pre-processing step 200a. First, step 210a is implemented to select a set of feature matrices after pre-modeling feature engineering from the hypothesis models pre-stored in the intelligent pre-diagnosis and health management system. Next, the monitoring data obtained in step 200a is converted into a feature matrix of the same format as the feature matrix after the feature engineering is performed before the modeling (ie, step 220a).

For the pre-processing step 200b, please refer to "Fig. 2B". First, step 210b is implemented to perform the data cleaning step as described above. Next, a statistical feature processing step 220b and a blending feature processing step 230b are implemented. Finally, implementation step 240b places all features in the feature pool as the basis for the lead assist parameter analysis step 300b in the next phase.

Next, the similarity analysis step 300a of the branch 1 and the leading auxiliary parameter analysis step 300b of the branch 2 will be described.

Please refer to FIG. 3A, which is a flowchart of the similarity analysis step 300a of the branch 1 according to an embodiment of the present invention. The purpose of the similarity analysis step 300a of the branch 1 is to understand whether there is a hypothesis model suitable for analyzing the new machine equipment in the hypothesis model pre-stored in the intelligent pre-diagnosis and health management system of the present invention (step 330a) . Through the above similarity analysis, if the pre-existing hypothesis model has a hypothesis model, the feature set before modeling is similar to the characteristic matrix of the monitoring data node of the analysis tree node (Object), which is not only higher than the specified threshold value. And the spatial distance is also the closest, indicating that the hypothesis model is suitable for the prediction of the analysis tree node (Object) specific to the new machine (step 350a). Examples of quantization methods suitable for the similarity analysis step 300a may include Eucliled Distance, Mahalanobis Distance, Manhattan Distance, Minkowski distance (Minkowski distance) ), Cosine Similarity, etc. The functions and types are well known to those skilled in the art and will not be described here.

In another case, if the similarity comparison result is lower than the specified threshold, there is no hypothesis model applicable to the new machine among the built-in reference hypothesis models. At this time, the feature matrix composed of the key parameter (CP) and the auxiliary parameter (AP) is obtained as described in the branch 2 step 340a and the step 360a (step 340a), and the supervised learning hypothesis method is specified (step 360a) is used as the basis for modeling the branch 1 predictive modeling step 400a.

After the step 400a is completed, the modeled hypothesis model indicator of the analysis tree node may be added or updated to the mapping table of the intelligent prediction and health management object analysis tree module 30 and the analysis tree is completed. The branch 1 of the node predicts the hypothesis model expansion or update.

Please refer to the reference figure "3B", which is a flowchart of the leading auxiliary parameter analysis step 300b of the branch 2 of an embodiment of the present invention. The lead assist parameter analysis step 300b includes a set key parameter step 310b, a set feature parameter step 320b, a feature selection step 330b, a causal analysis step 340b, an auxiliary parameter setting step 350b, and a lead assist parameter step 360b.

First, the setting key parameter step 310b and the setting characteristic parameter step 320b are performed after the monitoring data of the analysis tree node is cleared by a domain expert, and the key parameter (CP) is specified by a domain expert, and then the feature is captured through the feature. The method converts the data into statistical features and mixed features (such as steps 220b and 230b of FIG. 2B), and other parameters other than the key parameters (CP) form a parameter that is closest to the key parameter (CP).

These features form a large set of features and classify the feature set into a set of key parameters (CP) and a set of features without key parameter (CP) features. Next, in the causal analysis step (step 340b), the effective leading auxiliary parameter (LAP) of the key parameter (CP) is found through the Granger Causality Test.

If the feature set is too large, the feature selection step may be changed (step 330b), and the feature with the highest correlation with the key parameter (CP) is selected first, and then the Granger causality check is performed, and the analysis tree is used as the analysis tree. The branch 2 of the node predicts the basis of the hypothesis model modeling step (step 400b).

After the step 400b is completed, the hypothesis model indicator of the analysis tree node (Object) can be added or updated to the mapping table of the intelligent prediction and health management object analysis tree module 30 and the analysis tree node is completed. The branch 2 hypothesis model of (Object) is expanded or updated.

Regarding the analysis flow of the above-described Granger causality check step 400b and the branch 2 hypothesis model modeling step 400b, first, it is assumed that the key parameter (CP) and the selected one auxiliary parameter (AP) are a stationary time series. . The null hypothesis is: "The auxiliary parameter is not the Granger cause of the key parameter", and an autoregressive model (AR model on CP) of the key parameter is established, as shown in the following equation 1:

[Formula 1]

Where CP _t represents the CP observation at time t, according to the F test, if the backward period CP _t is added to the model, the explanatory power of the autoregressive model can be improved, and the backward period will be left in the model, and m represents the key parameter. The variable lag period is determined to be the most significant one at the time, and error _t is the estimated error term.

Add the backward period of the auxiliary parameter (AP) and establish the model as Equation 2:

[Formula 2]

Similarly, according to the F-check, when the auxiliary parameter is added to the model to increase the explanatory power of the model, the backward period will be left in the model. Where p is the earliest time in the backward period of the auxiliary parameter (AP) variable, and q is the most recent time in the backward period of the auxiliary parameter (AP) variable.

If no backward period of the auxiliary parameter (AP) is left in the model, the null hypothesis without Granger causality is established.

If the auxiliary parameter (AP) has a causal relationship with the key parameter (CP), the auxiliary parameter (AP) is included in an associated parameter candidate set.

Finally, all the auxiliary parameters (AP) of the auxiliary parameter candidate set are subjected to another F-test in the following two models (Formula 3, Equation 4) to confirm how long the auxiliary parameter (AP) can be earlier than the key parameter (CP). The change produces a reaction. Compared with Equation 4, Equation 3 has an additional period of information AP _{t -q} . Therefore, when comparing the results of Equations 3 and 4, it can be judged whether there is a difference in the data of the period, and if so, it represents More information on this period is available.

[Formula 3]

[Formula 4]

Finally, we select the auxiliary parameter (AP) that can reflect this key parameter (CP) change as a leading auxiliary parameter (LAP).

Refer to Figure 4 for an example of the model adaptive optimization step (adaptability switching mechanism, step 500) mentioned in Figure 1B.

First, step 510 is implemented to confirm whether a machine cannot continue to transport. If the result of "determining whether the machine is unable to continue the transportation" is "Yes", then step 520 is implemented to obtain the monitoring data and the maintenance record of the analysis tree node (Object) currently under the monitoring point. Among them, the maintenance records correspond to each monitoring data in a one-to-one manner. Each of the analysis tree nodes has at least one monitoring point, and each maintenance record yi includes a maintenance status value of at least one monitoring point. In a specific example, an analysis tree node can represent an organic metal. A metal-organic chemical vapor deposition (MOCVD) machine, and the machine includes k monitoring points, that is, one analysis tree node (device) contains at least one monitoring such as current, voltage, and vibration frequency. point.

Next, step 530 is performed to calculate the model evaluation index (MEI) of the branch 1 and the branch 2. The common model evaluation indicators include, but are not limited to, the accuracy rate, the recall rate, the F value, and the area under the ROC curve.

Subsequently, steps 540 and 550 are implemented to confirm whether the model evaluation index of any branch prediction model exists in the bi-branch modeling mode of the analysis tree object is better than the other branch model. Specifically, the above steps are all related to the calculus, and in step 540, according to the above-mentioned model evaluation index (MEI) result, it is determined whether the subsequent step 550 uses the best solution to inherit the past training experience (the preset is optimal in this embodiment). The scheme is branch 2).

If the result of the analysis is that the MEI of the hypothesis model is better than the branch 2 model for m consecutive times, then step 560 is implemented, that is, the branch 1 prediction model is set as the golden model of the analysis tree and the system is started. Pre-diagnostic analysis; but conversely, if the model evaluation index of the branch 1 hypothesis model is not superior to the branch 2 model for m consecutive times, then step 570 is implemented and the branch 2 hypothesis model is specified as the best solution and the pre-diagnostic analysis of the system is started. Converting the analysis tree object monitoring data into the same feature format as step 300a, finally reconstructing the prediction model of branch 1 and adding or updating the hypothesis model indicator of the analysis tree object to the intelligent prediction and The mapping table in the health management object analysis tree module 30 is used to complete the expansion or update of the branch 1 hypothesis model.

The above embodiments can be implemented by using a computer program product, more specifically, as a computer program product and by performing the steps in the above embodiments through a system readable medium including a plurality of instructions.

The system readable medium including a plurality of instructions includes, but is not limited to, a floppy disk, a compact disk, a CD-ROM, a magneto-optical disk, a read-only memory, a random access memory, and an erasable programmable read-only memory ( EPROM), electronic erasable programmable read only memory (EEPROM), optical card or magnetic card, flash memory, or any machine readable medium suitable for storing electronic instructions. Alternatively, the intelligent pre-diagnosis and health management system 10 described above can also be downloaded as a computer program product and transferred from a remote location to a local computer via a data signal such as a network connection.

The invention can effectively utilize the two-branch modeling mode, and quickly start predictive analysis for a new machine with only a small amount and no maintenance record. And as the monitoring data and the corresponding maintenance records continue to increase, the intelligent pre-diagnosis and health management system of the present invention branches to "Yes" after the "cannot continue the maintenance" of each analysis tree node. 1 and branch 2 prediction hypothesis model comparison, when it is confirmed that there is a branch prediction model in the bifurcation model, the model evaluation index is better than the other branch prediction model for m times, that is, the better branch prediction model is used as the analysis. The golden node of the tree node and the best decision make the prediction results of the intelligent pre-diagnosis and health management system meet the expected target value.

The present invention has been described in detail above, but the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the scope of the invention. That is, the equivalent changes and modifications made by the scope of the present application should remain within the scope of the patent of the present invention.

10‧‧‧Smart pre-diagnosis and health management system

20‧‧‧Analysis Engine Service Management Module

30‧‧‧Intelligent Forecasting and Health Management Object Analysis Tree Module

40‧‧‧Machine Learning Library Module

50‧‧‧File System Module

60a‧‧‧First exchangeable application interface

60b‧‧‧Second exchangeable application interface

60c‧‧‧ Exchangeable Driver Interface

70‧‧‧External Machine Learning Module

80‧‧‧External Reference Model Module

90‧‧‧External data collection drive

91‧‧‧External database

110, 120, 200a, 210a, 220a, 200b, 210b, 220b, 230b, 240b, 300a, 300b, 310a, 310b, 320a, 320b, 330a, 330b, 340a, 340b, 350a, 350b, 360a, 360b, 400a, 400b, 500, 510, 520, 530, 540, 550, 560, 570 ‧ ‧ steps

FIG. 1A is a structural diagram of a smart pre-diagnosis system according to an embodiment of the present invention. FIG. 1B is a flowchart of a method for modeling a smart pre-diagnosis system according to an embodiment of the present invention. 2A to 2B are flowcharts of the data pre-processing step a and the data pre-processing step b according to an embodiment of the present invention. 3A to 3B are flowcharts of the similarity analysis step and the lead assistant parameter analysis step according to an embodiment of the present invention. FIG. 4 is a flowchart of a model adaptive optimization step according to an embodiment of the present invention.

Claims (11)

A smart pre-diagnosis and health management system modeling method is implemented by a smart pre-diagnosis and health management system with a plurality of reference hypothesis models built therein, the method comprising: a new tree establishment step, which is based on a to-be-monitored The machine defines at least one analysis tree node, and each analysis tree node acquires a monitoring data via at least one monitoring point; the double branch modeling step is performed on a branch 1 to implement a data preprocessing step a Converting the monitoring data into a specified feature format and performing a similarity analysis to select a reference hypothesis model with the highest similarity and exceeding a specified threshold from the reference hypothesis models as a branch 1 prediction hypothesis model of the analysis tree node Simultaneously performing a data pre-processing step b on the monitoring data in a branch 2 to confirm and construct a key parameter (CP) corresponding to the analysis tree node (Object) and a plurality of related parameters (AP) by using a causal relationship check a hypothesis model applicable to the machine to be monitored as a branch 2 prediction hypothesis model of the analysis tree node; and an excellent model self-adaptation In the step, as the monitoring data is continuously generated, after the analysis tree node (Object) performs a judgment that "whether the machine can not continue the maintenance", if the determination result is "Yes", the double branch is constructed. Selecting an optimal solution among the branch 1 prediction hypothesis model or the branch 2 prediction hypothesis model constructed by the modular step as a benchmark for optimizing the decision of the analysis tree node, so that a prediction result of the system conforms to the expectation Target value. The modeling method of claim 1, wherein in the new tree establishing step, the machine to be monitored is a new machine with no maintenance record. The modeling method of claim 1, wherein in the branch 1 of the two-branch modeling step, the specified feature format has the same feature format as before the reference hypothesis model is modeled. The modeling method of claim 3, wherein after the model optimization step specifies the optimal solution, the method further includes converting the monitoring data into the specified feature format and updating to the smart pre-diagnosis and These reference hypothesis models built into the health management system. The modeling method of claim 1, wherein in the model optimization step, the method further includes acquiring at least one maintenance record, the group maintenance record including at least one maintenance status value of the monitoring point, so that the monitoring The information is in a one-to-one relationship with the maintenance record. The modeling method of claim 1, wherein in the branch 1 of the two-branch modeling step, the similarity analysis adopts at least one selected from the Eucliled Distance, Maha The Mahalanobis Distance, the Manhattan Distance, the Minkowski distance, the Cosine Similarity, and combinations thereof are used as a quantification method for the similarity analysis. The modeling method of claim 1, wherein in the branch 1 of the two-branch modeling step, when the result of the similarity analysis does not exceed the specified threshold, the branch 2 is first The identified critical parameter (CP) and the characteristic matrix composed of the relevant parameters (AP) are used as the basis for modeling the branch 1 prediction hypothesis model, and after obtaining a maintenance record label in the model adaptive optimization step, The supervised learning method reconstructs the branch 1 prediction hypothesis model. The modeling method of claim 7, wherein the supervised learning method comprises at least one selected from the group consisting of a support vector machine (SVM), a regression analysis (Regression), a random forest (Random Forest), A group consisting of its combination. The modeling method according to claim 1, wherein in the model adaptive optimization step, when the branch 1 prediction hypothesis model is superior to the branch 2 prediction hypothesis model for m consecutive times, the branch 1 prediction is set. The hypothesis model is the best solution; when the branch 1 prediction hypothesis model is not better than the branch 2 prediction hypothesis model for m consecutive times, the branch 2 prediction hypothesis model is designated as the optimal solution. The modeling method of claim 9, wherein in the model adaptive optimization step, the m value is a positive integer greater than two. A computer program product for a smart pre-diagnosis and health management system, which enables the intelligent pre-diagnosis and health management system modeling method as described in any one of claims 1 to 10 when the computer program product is executed .