KR20200003323A - Analysis method of controller for factory automation using artificial intelligence-based reverse engineering - Google Patents

Abstract

The present invention relates to an analysis method of a factory automation controller using artificial intelligence-based reverse engineering, which comprises: a) an input step of receiving and preprocessing input and output data of an analysis target intelligent controller; b) a step of constructing a logical model of the preprocessed input and output data through artificial intelligence learning; and c) a step of extracting internal logic of the analysis target intelligent controller through the constructed model.

Description

Analysis method of controller for factory automation using artificial intelligence-based reverse engineering}

The present invention relates to an analysis method of a factory automation controller using artificial intelligence-based reverse engineering, and more specifically, a programmable logic controller (PLC), a direct coefficient control device (Direct Digital) using a machine learning algorithm. The present invention relates to a method for identifying and reproducing the structure and operation principle of an intelligent controller such as a controller (DDC), a distributed control system (DCS), a standalone controller, and the like.

Recently, intelligent controllers have been actively introduced for factory automation such as smart factories. The intelligent controller is a concept that includes a programmable logic controller (PLC), a direct digital controller (DDC), a distributed control system (DCS), and a standalone controller. .

For factory automation, it is very important to understand the coupling of the organic operation of the above controllers. However, up to now, in Korea, when constructing various process control systems, equipment is imported from overseas manufacturers, and since the operation principle and organic coupling relationship of imported equipment cannot be analyzed, it is expensive to maintain the equipment. have.

"Abdelhameed, Magdy M., Darabi, Houshang. Diagnosis and debugging of programmable logic controller control programs by neural networks.Proceedings of the 2005 IEEE International Conference on Automation Science and Engineering, Edmonton, Canada, August 1 & 2, 2005, ISBN: 0-7803-9426-7, pp. 313-318, presents the methods necessary to model programmable logic controllers using cyclic neural networks and to detect or diagnose errors.

However, due to the lack of tracking and analysis of algorithm learning processes, the internal working principle of the controller cannot be identified, and there is a limit that it is difficult to verify model accuracy because various algorithms are not applied.

hofer, Herbert; Schatz, Roland; Wirth, Christian; Mossenbock, Hanspeter. A Comprehensive Solution for Deterministic Replay Debugging of SoftPLC Applications. IEEE Transactions on Industrial Informatics, Volume 7, Issue 4, November 2011, Article number 6009197, ISSN: 15513203, 641-651쪽”, “Pr

hofer, Herbert; Wirth, Christian; Berger, Richard. Reverse engineering and visualization of the reactive behavior of PLC applications. IEEE International Conference on Industrial Informatics (INDIN), 2013, Article number 6622946, ISBN: 978-1-4799-0752-6, 564-571쪽”, “Pirzadeh, Heidar; Agarwal, Akanksha; Hamou-Lhadj, Abdelwahab. An Approach for Detecting Execution Phases of a System for the Purpose of Program Comprehension. 2010 8th ACIS International Conference on Software Engineering Research, Management and Applications, Article number 5489832, ISBN: 978-076954075-7, 207-214쪽”에서는 프로그램 가능한 로직 컨트롤 시스템의 추적 로그(trace log)를 이용하여 콘트롤러를 모델링하고 오류를 검출하는 방법을 제시하였다.“Pr

hofer, Herbert; Schatz, Roland; Wirth, Christian; Mossenbock, Hanspeter. A Comprehensive Solution for Deterministic Replay Debugging of SoftPLC Applications. IEEE Transactions on Industrial Informatics, Volume 7, Issue 4, November 2011, Article number 6009197, ISSN: 15513203, pp. 641-651 ”,“ Pr

hofer, Herbert; Wirth, Christian; Berger, Richard. Reverse engineering and visualization of the reactive behavior of PLC applications. IEEE International Conference on Industrial Informatics (INDIN), 2013, Article number 6622946, ISBN: 978-1-4799-0752-6, pp. 564-571 ”,“ Pirzadeh, Heidar; Agarwal, Akanksha; Hamou-Lhadj, Abdelwahab. An Approach for Detecting Execution Phases of a System for the Purpose of Program Comprehension. In the 2010 8th ACIS International Conference on Software Engineering Research, Management and Applications, Article number 5489832, ISBN: 978-076954075-7, pages 207-214, modeling the controller using the trace log of a programmable logic control system. And a method of detecting an error is presented.

However, such a conventional method has a big limitation in that it is modeled with a certain degree of internal system operation, and in particular, it is impossible to apply when trace log information cannot be obtained. .

Prhofer, Herbert; Schatz, Roland; Wirth, Christian; Mossenbock, Hanspeter. A Comprehensive Solution for Deterministic Replay Debugging of SoftPLC Applications. IEEE Transactions on Industrial Informatics, Volume 7, Issue 4, November 2011, Article number 6009197, ISSN: 15513203, 641-651쪽

Prhofer, Herbert; Wirth, Christian; Berger, Richard. Reverse engineering and visualization of the reactive behavior of PLC applications. IEEE International Conference on Industrial Informatics (INDIN), 2013, Article number 6622946, ISBN: 978-1-4799-0752-6, 564-571쪽 Pirzadeh, Heidar; Agarwal, Akanksha; Hamou-Lhadj, Abdelwahab. An Approach for Detecting Execution Phases of a System for the Purpose of Program Comprehension. 2010 8th ACIS International Conference on Software Engineering Research, Management and Applications, Article number 5489832, ISBN: 978-076954075-7, 207-214쪽 Abdelhameed, Magdy M., Darabi, Houshang. Diagnosis and debugging of programmable logic controller control programs by neural networks. Proceedings of the 2005 IEEE International Conference on Automation Science and Engineering, Edmonton, Canada, August 1 & 2, 2005, ISBN: 0-7803-9426-7, pp. 313-318

Prhofer, Herbert; Schatz, Roland; Wirth, Christian; Mossenbock, Hanspeter. A Comprehensive Solution for Deterministic Replay Debugging of SoftPLC Applications. IEEE Transactions on Industrial Informatics, Volume 7, Issue 4, November 2011, Article number 6009197, ISSN: 15513203, 641-651

Prhofer, Herbert; Wirth, Christian; Berger, Richard. Reverse engineering and visualization of the reactive behavior of PLC applications. IEEE International Conference on Industrial Informatics (INDIN), 2013, Article number 6622946, ISBN: 978-1-4799-0752-6, pp. 564-571 Pirzadeh, Heidar; Agarwal, Akanksha; Hamou-Lhadj, Abdelwahab. An Approach for Detecting Execution Phases of a System for the Purpose of Program Comprehension. 2010 8th ACIS International Conference on Software Engineering Research, Management and Applications, Article number 5489832, ISBN: 978-076954075-7, pp. 207-214

The technical problem to be solved by the present invention is to operate the analysis target to obtain input and output data, and process it through a machine learning algorithm to build a decision model, based on artificial intelligence that can grasp the operating principle inside the control system in reverse An analysis method of a factory automation controller using reverse engineering is provided.

Another technical problem to be solved by the present invention is to classify according to whether the output value of the intelligent controller is discrete or continuous value, and by providing a suitable algorithm for each to build a high accuracy machine learning model, the learning of the algorithm after The present invention provides an analysis method of a factory automation controller using artificial intelligence-based reverse engineering that can track the process, grasp the structure and operation principle of an intelligent controller, and further model an alternative controller.

The analysis method of the factory automation controller using the artificial intelligence-based reverse engineering to solve the above problems, a) receiving input and output data of the analysis target intelligent controller, and b) pre-processed input and output data Constructing a logic model through artificial intelligence learning; and c) extracting internal logic of an intelligent controller to be analyzed through the constructed model.

According to an embodiment of the present invention, the input and output data in step a) is discrete or continuous, and the learning model of step b) is different depending on the type of data, and the preprocessing is performed. Discretization can be performed in units of arrays.

According to an embodiment of the present invention, the step b) includes performing a decision tree classifier learning, checking whether the model accuracy is greater than or equal to the set criterion, and performing the step a) if the model accuracy is greater than or equal to the predetermined criterion. And c) performing step c).

According to an embodiment of the present invention, step b) may be performed using a neural network model when the input / output data is discrete.

According to an embodiment of the present invention, step c) may create a flowchart of a career decision process after analysis using a tree classifier, and create a ladder logic diagram based on the flowchart to visualize the career decision. .

According to an embodiment of the present invention, the step c) may extract the internal logic of the intelligent control device using a neural network model or a linear regression model.

The analysis method of the factory automation controller using artificial intelligence-based reverse engineering of the present invention, by using the input and output data of the intelligent controller can grasp the operating principle of the control system, it is possible to secure the advanced process control technology.

The present invention can localize the intelligent controllers when new plant construction or expansion of the same process, there is an effect that can be easy to maintain and repair.

1 is a flowchart of an analysis method of a controller for factory automation using artificial intelligence-based reverse engineering according to a preferred embodiment of the present invention.
2 is a flowchart of the algorithm of FIG.
3 to 6 are exemplary diagrams of histograms of input variables P1 to P4 respectively input to the programmable logic controller PLC that is an analysis target.
7 to 12 are exemplary diagrams of histograms of output variables ACC, DEC, STD, FORWARD, RUN, and STOP.
13 is a flowchart of a career decision process.
14 to 19 are exemplary diagrams of ladder logic diagrams of output variables created by analyzing tree-based models, respectively.
20 is an exemplary diagram of a neural network model selected in a model generation process of an output variable ACC.
21 is an exemplary diagram of a neural network model selected in the model generation process of the output variable DEC.
22 is an exemplary diagram of a ladder logic diagram of an output variable DEC prepared by analyzing a neural network model.

Hereinafter, an analysis method of a factory automation controller using artificial intelligence-based reverse engineering of the present invention will be described in detail with reference to the accompanying drawings.

Embodiments of the present invention are provided to more fully describe the present invention to those skilled in the art, and the embodiments described below may be modified in various other forms, and The scope is not limited to the following embodiments. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the inventive concept to those skilled in the art.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms "a", "an" and "the" may include the plural forms as well, unless the context clearly indicates otherwise. Also, as used herein, "comprise" and / or "comprising" specifies the presence of the mentioned shapes, numbers, steps, actions, members, elements and / or groups of these. It is not intended to exclude the presence or the addition of one or more other shapes, numbers, acts, members, elements and / or groups. As used herein, the term "and / or" includes any and all combinations of one or more of the listed items.

Although the terms first, second, etc. are used herein to describe various members, regions, and / or portions, it is obvious that these members, components, regions, layers, and / or portions are not limited by these terms. . These terms do not imply any particular order, up or down, or superiority, and are only used to distinguish one member, region or region from another member, region or region. Accordingly, the first member, region, or region described below may refer to the second member, region, or region without departing from the teachings of the present invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings schematically showing embodiments of the present invention. In the drawings, for example, variations in the shape shown may be expected, depending on manufacturing techniques and / or tolerances. Accordingly, embodiments of the present invention should not be construed as limited to the specific shapes of the regions shown herein, but should include, for example, changes in shape resulting from manufacturing.

1 is a flowchart of an analysis method of a controller for factory automation using artificial intelligence-based reverse engineering according to a preferred embodiment of the present invention, and FIG. 2 is an algorithm flowchart.

1 and 2, the analysis method of the factory automation controller using artificial intelligence-based reverse engineering according to a preferred embodiment of the present invention, input step (S10), generation step (S20), extraction step (S30) , Analysis step (S40). The analysis method of the factory automation controller using the artificial intelligence-based reverse engineering of the present invention is made using a device including a controller capable of a conventional operation and a storage device and a memory capable of temporarily storing data, for example, a personal computer. Can be used.

First, the input step (S10) receives the input and output data from the sensors mounted on the intelligent controller to be analyzed and preprocessing.

Next, the generation step (S20) selects an appropriate prediction model according to the type of data input in the input step, and then trains.

The type of input data may be divided into discrete or continuous values.

In the extraction step (S30) to find the internal logic of the intelligent control device by analyzing the model learned in the generation step (S20), a tree-based model, a neural network model or a linear regression model may be used.

Next, in the verification step (S40), the logic extracted in the extraction step (S30) is verified.

Hereinafter, each step of the present invention will be described in more detail.

First, in the input step (S10), the sensor input and output data of the intelligent controller is received and preprocessed. At this time, since the logic analysis method of the intelligent control device varies according to the type of input / output data, the data is preprocessed.

In the input step S10, the input / output data in the form of a continuous time series is discretized and stored in an array form at a minimum time interval, for example, in microseconds, without losing information on the change in values. .

If the prediction is sufficient for input at the present time, the sequence is used without further data preprocessing. On the other hand, when building a predictive model requires historical data, for example, if it includes a time delay module, when applying a tree-based model, it determines how old the data is to be used and a specific time step. Truncates the inputs by the amount of time to create a time window.

In the case where the logic using the output values of the immediately preceding time point as the input value of the present time is frequently displayed, for example, a self-holding logic unit is included, the immediately preceding output values may also be classified ( classifier). In the case of outputs in the form of step functions rather than pulses, it is more intuitive to predict when the signal is turned on and off, so preprocess the data accordingly.

3 to 6 are exemplary diagrams of histograms of the input variables P1 to P4 input to the programmable logic controller PLC as an analysis target, and FIGS. 7 to 12 are output variables ACC, DEC, STD, FORWARD, An illustration of the histogram of RUN and STOP).

In the generation step S20, when the data type input in the input step S10 is a discrete value, learning is performed by using a tree-based model.

In the present invention, the decision tree classifier is learned while adjusting the time interval. You can use the ROC AUC score as an indicator of prediction. It is important to make the tree decision model as simple as possible to find the logic of the intelligent controller to be analyzed from the learned tree. Therefore, the time interval of the data used as the input value is minimized and the model of the tree is made as simple as possible.

Note that in the process of model construction, if an inappropriate metric is used as an evaluation indicator, the score may be high even if the prediction is completely out of order. For example, simply using square error as a quantitative function yields a high score even if all predictions are zero for the output of the beat form.

The generating step S20 includes learning a decision tree classifier S21 and determining whether the model accuracy is greater than or equal to a reference value S22.

After performing this generation step (S20), in the extraction step (S30) to obtain the desired accuracy by learning the decision tree model and then visualize the decision path of the model to analyze the logic of the controller to be analyzed do.

The logic built in the above process is only valid for the input pattern used for learning. In particular, the use of past data as an input value increases the number of possible patterns. It is not practical to generate output data and train the model for every pattern, but try to learn as many different time series patterns of the input as possible.

In addition, when the form of the output data is a discrete value, the generation step (S20) may additionally select a neural network model when selecting an algorithm. In the present invention, a hard sigmoid function may be used as an activation function in a neural network model learning method, but is not necessarily limited to this function.

In the generation step (S20), the neural network model is trained while adjusting the time interval. Use ROC AUC scores as an indicator of forecasting. The neural network model can be configured in various ways according to the situation. Two models were used in the present invention. First, when it is sufficiently predictable with only the input data of the present time, a neural network model composed of a fully connected layer may be used. Next, when data from the past point in time is required, a neural network model including a locally connected layer may be selected and used.

In this case, the weight of the neural network model learned in the extraction step (S30) is simplified in a range in which accuracy is maintained so that translation is easy to the controller logic to be analyzed. Next, draw the ladder logic diagram by analyzing the weights obtained.

If it's too difficult to draw a diagram, the logic is represented in formula form.

As a result of analyzing the output variable 'FORWARD' illustrated in FIG. 10, a career decision process obtained through learning a decision tree model may be represented as in the flowchart of FIG. 13.

As an example of the extraction step (S30) using the tree model, the result of FIG. 13 is represented by a ladder logic diagram, as shown in FIG. 14.

The career decision process of FIG. 13 is analyzed by following the top-down method. Binary classification at each node means on-off of the logic unit indicated at the node. For example, in FIG. 13, when the uppermost node is interpreted while following the left child node, when FORWARD (t-1) is on and P1 (t) is off, the current state of FORWARD (t) is on. Meanwhile, following the right child node, if FORWARD (t-1) is off and P4 (t) is on, FORWARD (t) is on.

This analysis requires that FORWARD (t-1) | / | beneath the OR circuit. After the verification step (S40), even if this logic unit was excluded, it was confirmed that the ROC AUC score was maintained as it was, and the final result was excluded. This process makes it possible to modify the actual PLC device with simpler and more intuitive logic.

The results of creating the ladder logic diagram through the analysis of the output variables 'ACC', 'DEC', 'RUN', 'STD', and 'STOP' are shown in FIGS. 15 to 19, respectively.

In neural network models, pay attention to the problems of inadequate metric, overfitting, etc. In the case of the experimental data in the present invention, one layer is sufficient, but in the case of complex logic used in the actual field, several layers may be required.

In one embodiment of the present invention, the decision tree model is first implemented, and the model candidate group, such as how to stack the hierarchy based on the result, the weight selection, and the pruning process, may be selected. The proper use of the tree decision model as a preceding model can reduce trial and error in the neural network model construction process.

It is assumed that most of the weights of the local link layers used when learning historical data together can be deleted through pruning. In fact, how PLCs work in this form of modeling is possible because most of the historical data, except for some, does not affect the current results. This also makes the logic more understandable. The pruning process may be performed by removing weights having the smallest weight one by one at specific intervals in a specific epoch section in the learning process.

The following is an example of the extraction step (S30) when using the neural network model. In the case of the output variable 'ACC', logic prediction can be performed without the need for data from the past. In this case, the analysis is sufficient with a simple neural network model as shown in FIG. 20.

Meanwhile, in the case of the output variable 'STD', sufficient accuracy may be obtained by learning data of a past time point together. In this case, the local connection layer as shown in FIG. 21 is used to make the weight analysis easy, and the vestibular process described above is included in the model training. The weight obtained by learning each model is represented by Equation 1 below.

If h ₁ and h ₂ are defined in Equation 2, Equation 2 below.

The activation function σ is a hard sigmoid function and may be defined by Equation 3 below.

Therefore, when at least one of P ₁ (t) and P ₄ (t) is 1, ACC (t) becomes 1, and when both are 0, the value of ACC (t) becomes 0. That is, the output variable ACC (t) corresponds to an OR gate of the input variables P ₁ (t) and P ₄ (t), which is represented by ladder logic in FIG. 15.

In a similar manner, a weight analysis of DEC (t) may be performed to obtain FIG. 22. Note that the input variables P ₁ and P ₄ contain the time delay module. 16 is ladder logic for the output variable 'DEC' obtained using the tree model. In other words, different ladder logic can be extracted by applying different prediction models to the same variables. This does not mean that each prediction is wrong, but that the two models predict equivalent results for a given pattern of input variables.

In the same vein, other ladder logic diagrams extracted in the present invention may not exactly match the structure of the actual intelligent controller, but no error occurs in the value of the predicted output variable.

In the generation step S20, when the output data has a continuous value, the algorithm selects one of linear regression or decision treeclassifier using various basic functions.

If you select a polynomial fitting model using a polynomial function, apply the linear regression model with increasing order. Raise the order by the desired accuracy. When a decision tree classifier is selected, when the variation in the resulting values, y (t)-y (t-1), is categorical or can be rounded appropriately to make it categorical We can turn the regression problem into a classification problem.

As a way of interpreting the results, we check the derived equations when applying the linear regression algorithm. If the decision tree classifier is applied, the output data is regarded as a discrete value and goes through the same process as in the first case.

As such, the internal logic of the PLC can be checked through the extraction step (S30). In the verification step (S40), the input value is input again to the constructed logic, and the output is compared to verify whether the logic is correct. If necessary, more accurate logic can be obtained by excluding and verifying unnecessary logic units at the user's discretion.

It will be apparent to those skilled in the art that the present invention is not limited to the above embodiments and may be variously modified and modified without departing from the technical spirit of the present invention. will be.

Claims (6)

a) pre-processing input / output data of the analysis target intelligent controller;
b) constructing a logical model of the preprocessed input / output data through artificial intelligence learning; And
c) Analysis method of a factory automation controller using artificial intelligence-based reverse engineering comprising extracting the internal logic of the intelligent controller to be analyzed through the built model. The method of claim 1,
The input / output data in step a) is discrete or continuous, and the learning model of step b) is different depending on the type of data.
The pre-processing, the analysis method of the factory automation controller using artificial intelligence-based reverse engineering, characterized in that the input-output data Discretization (microsecond) unit to process in the form of an array (array). The method of claim 2,
B),
Performing decision tree classifier learning; And
Check if the model accuracy is above the set criteria, and if it is not above the set criteria, repeat step a).
Method of analysis of a factory automation controller using artificial intelligence-based reverse engineering comprising the step of performing the step c) if the reference. The method of claim 2,
B),
If the input and output data is discrete, the analysis method of the factory automation controller using artificial intelligence-based reverse engineering, characterized in that learning using a neural network model. The method of claim 1,
C),
Using the tree classifier, make a flow chart of the career decision process after analysis.
Create a ladder logic diagram based on the flowchart above,
An analysis method of a factory automation controller using artificial intelligence-based reverse engineering characterized in that to visualize the career decision. The method of claim 1,
C),
An analysis method for a factory automation controller using artificial intelligence-based reverse engineering, which extracts internal logic of an intelligent control device using a neural network model or a linear regression model.

Images