KR102221473B1 - System for error detection according to plc control - Google Patents

Abstract

The error detection system according to PLC control of the present invention uses the state values at the same time for each of a plurality of device operations according to the PLC input/output signal of the PLC (Programmable Logic Controller) control system as state transition information for the PLC control system. A first error detector configured to detect an error in the PLC control system by detecting and comparing the detected state transition information with reference state transition information during normal operation of the PLC control system; And dividing a power consumption signal according to the control of the PLC control system or a PLC input/output signal according to the control of the PLC control system, extracting a plurality of statistical feature information from the divided signal, and training the extracted statistical feature information. It characterized in that it comprises a second error detection unit for detecting the abnormality of the PLC control system by applying to the self-associative neural network (AANN).

Description

Error detection system according to PLC control {SYSTEM FOR ERROR DETECTION ACCORDING TO PLC CONTROL}

The present invention relates to a programmable logic controller (PLC) control, and more particularly, provides a technology for an automation tool capable of accurately detecting and quarantining defects and abnormalities in a PLC control manufacturing system.

PLC is an industrial computer programmed in a low level language and used to control automation systems. The PLC program, the internal logic of PLC, controls the automation system through Boolean operation. The PLC program designed in the general manufacturing process goes through a verification process, and if the accuracy of the program is guaranteed, the actual automation system is controlled.

In recent years, in the automated manufacturing industry, as the complexity of the manufacturing line increases, the control logic is vast and has a very complex design. Accordingly, the PLC program is also complexly logicd. For this reason, diagnosis and monitoring of PLC programs is also becoming more and more difficult, and accordingly, the time taken to detect and correct errors is gradually increasing. According to Operational Diagnostics, The holy grail of control automation, delays due to these diagnostic methods and the time it takes to identify errors exceed 80% of the total plant failure time. In particular, in the case of an automobile body assembly line, the average cycle time is around 1 minute, and thus, when the line is stopped due to equipment failure, it causes a large loss of profit for a short period of time.

A typical automation system consists of a number of robots and automated transfer devices. The robot and the transfer device perform various tasks such as welding or transfer according to the logic of the PLC program. Recently, the automation system has built a large-scale automated production line, so it has high complexity and includes various factors of failure such as errors of the facility itself or errors caused by external factors such as interference of the robot's range of motion. Delays due to failures during operation cause enormous economic losses due to error detection and increased set-up time of the device.

In order to diagnose these errors, the automation system is monitored by adding diagnostic codes inside the PLC program that controls the flow of processes and logistics. In general, the automotive industry, a representative of automation systems, monitors the PLC program through a predicted abnormality display module when an error occurs in an automated manufacturing line. In other words, since the conventional monitoring method predicts an area with a high possibility of error occurrence and separately writes and adds a code according to the object for error diagnosis, all signals cannot be monitored. Therefore, the process to be monitored can be said to be very limited, and there is a limit as a monitoring method to detect errors that occur gradually. That is, it has a problem that it is difficult to respond in advance to the phenomenon of work failure occurring due to gradual wear of an apparatus or accessory.

In particular, in the case of a system abnormality that does not significantly affect the operation of the PLC control system, there is a problem in that it is not possible to detect all types of errors related to analog input/output signals because it is not possible to determine the abnormality through the control process model.

The present invention compares whether the state transition operation observed in the PLC control process matches the modeled state transition operation to detect whether an error in the PLC control system is detected, and the error in the PLC control system based on the PLC input/output signal or the power consumption signal. It relates to an error detection system according to PLC control that enables detection of whether or not.

The error detection system according to PLC control for solving the above problem is the state values for the PLC control system at the same time for each of a plurality of device operations according to the PLC input/output signal of the PLC (Programmable Logic Controller) control system. A first error detection unit detecting an error in the PLC control system by detecting as transition information and comparing the detected state transition information with reference state transition information during normal operation of the PLC control system; And dividing a power consumption signal according to the control of the PLC control system or a PLC input/output signal according to the control of the PLC control system, extracting a plurality of statistical feature information from the divided signal, and training the extracted statistical feature information. A second error detection unit applied to an autoassociative neural network (ANNN) to detect an abnormality in the PLC control system, and the first error detection unit includes a PLC digital input/output signal constituting the PLC input/output signal and A modeling module for modeling the reference state transition information for a PLC analog input/output signal; And a first comparing the reference state transition information modeled in the modeling module with the detected state transition information to determine whether an error exists in the PLC control system, and detecting an error type according to an error determination for the PLC control system. And an error detection module, wherein the modeling module performs MDSVTF modeling on the reference state transition information including transition timeout time information and signal state vector information corresponding to the PLC digital input/output signal.

delete

delete

The modeling module divides the PLC analog input/output signal into a plurality of segment segments based on a time point at which the PLC analog input/output signal rapidly fluctuates according to a statistical standard, and MDSVTF for the reference state transition information including quantization information of the segment segments. It is characterized by performing modeling.

The first error detection module is characterized in that when the detected PLC digital input/output signal and the on/off status of the reference state transition information according to MDSVTF modeling are different from each other, the detected PLC digital input/output signal is detected as an error. do.

The first error detection module is configured to detect the detected PLC analog input/output signal as an error when the detected PLC analog input/output signal and the signal state value of the reference state transition information according to MDSVTF modeling do not match. It is done.

The second error detection unit may include an input/output signal dividing module for dividing PLC input/output signals for operation of the PLC control system based on a power cycle; And a power signal dividing module for dividing the power consumption signal provided for the operation of the PLC control system for each device cluster.

The second error detection unit divides each of the divided power consumption signals or the divided analog input/output signals into three sections, and formulates an average value for each of the power consumption signals or the analog input/output signals subdivided into three sections. , Standard deviation formula, mean square formula, shape factor formula, peak value formula, peak-to-peak value formula, crest factor formula, impulse factor formula, redundant factor formula, distortion value formula, distortion factor formula, sharpness formula, sharpness factor formula And a statistical information detection module for extracting the statistical feature information by applying any one or more of the entropy equations.

The second error detection unit comprises a singular value detection module for extracting singular value information using a Singular Value Decomposition (SVD) algorithm from the divided power consumption signal or the divided analog input/output signal. It is done.

The second error detection unit performs a homogeneity mapping process using the self-associative neural network (ANNN) algorithm on the extracted statistical feature information and the singular value information to detect whether the PLC control system is abnormal. It characterized in that it comprises a module.

According to the present invention, it is possible to easily detect a state transition operation of a PLC control process, that is, a PLC control error using a DFA-based control process model, by using PLC control or recording log data of input/output signals.

In addition, it is possible to detect a PLC control error that is not detected by a PLC state transition operation without significantly affecting the system by using a process behavior model according to a PLC input/output signal or a power consumption signal.

1 is a block diagram of an embodiment for explaining an error detection system according to PLC control of the present invention.
2 is a reference diagram illustrating an MDSVTF control process model building procedure.
3 is a reference diagram illustrating an error and anomaly detection and isolation procedure of MPFDT.
4 is a reference diagram illustrating a scenario of a PLC control system for explaining the MDSVTF model building procedure.
5 is a timing diagram illustrating a signal state versus time chart of PLC input/output signals (sensor and actuator signals) of the PLC control system shown in FIG. 4.
6 is a reference diagram illustrating an automatic model of the PLC control system MDSVTF shown in FIG. 4.
7 is a reference diagram illustrating a stamping press system among PLC control systems.
8 is a reference diagram illustrating a state versus time chart of a PLC analog input/output signal in the stamping press system shown in FIG. 7.
9 is a reference diagram illustrating an MDSVTF model of the stamping press system shown in FIG. 7.
10 is a reference diagram illustrating an overall error detection procedure performed by a second error detection unit.
11 is a reference diagram of a working embodiment for explaining a process of dividing an analog input/output signal.
12 is a reference diagram illustrating a process of collecting power usage data.
13 is a reference diagram illustrating an operation process of a robot drilling subsystem, which relates to a small device cluster.
14 is a reference diagram illustrating a power consumption signal profile and a signal division process of the device cluster shown in FIG. 13.
FIG. 15 is a flowchart of an embodiment for explaining as a whole a procedure for statistical feature extraction, feature selection, and AANN learning performed by a second error detection unit.
16A to 16D are partial flowcharts for explaining the flowchart shown in FIG. 16 in detail.
17A to 17C are reference diagrams illustrating classification performance of a second error detection unit according to the present invention.

In general, detection and isolation of faults and abnormalities in PLC (Programmable Logic Controller) control systems is difficult. This invention proposes an automated tool for manufacturing process failure diagnosis called MPFDT (Manufacturing Process Failure Diagnosis Tool) that can accurately detect and isolate defects and abnormalities in PLC control systems.

The MPFDT according to the present invention uses two independent knowledge-based process operation models of the manufacturing system to meet the purpose of detection and isolation of defects and anomalies. Basically, it detects inconsistencies between the modeled process and the observed manufacturing process. The first process behavior model is error detection based on the DFA-based model of the PLC control process, which is used to determine whether the observed state transition behavior of the PLC control process matches the modeled state transition behavior. Log data records of PLC control or input/output signals are used to achieve a specific purpose. In most cases, when a device or part of a device fails, the state transition behavior of a normal PLC control process, i.e. MPFDT, can easily detect and isolate the MPFDT using a DFA-based control process model. This defect and anomaly search operation can be performed online.

On the other hand, all kinds of defects and abnormalities in the PLC control system (especially when they do not affect the state transition operation of the PLC control process much) cannot be detected using the control process model (the first process operation model). Accordingly, another independent and complementary knowledge-based process behavior model is provided.

The second process behavior model is basically used to identify one class classification criteria based on an autoassociative neural network (ANNN). It detects whether there is a significant difference between the observed power consumption profile of the PLC control system and the reference power consumption profile, which generally causes a large change in power consumption behavior due to the occurrence of a fault or anomaly, and affects the behavior of the control process. Crazy. Therefore, it is possible to detect errors by using power consumption signal analysis technology. In addition, it detects whether there is a significant change in the operation of the analog PLC input/output signal, and most errors and abnormalities related to the analog input/output signal can be detected using the second process behavior model.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of an embodiment for explaining an error detection system 10 according to PLC control of the present invention.

Referring to FIG. 1, an error detection system 10 according to PLC control includes a first error detection unit 100 and a second error detection unit 200.

The first error detection unit 100 detects status values at the same time for each of a plurality of device operations according to a PLC input/output signal of a PLC (Programmable Logic Controller) control system as status transition information for the PLC control system, The detected state transition information is compared with the reference state transition information during normal operation of the PLC control system to detect whether the PLC control system has an error.

MPFDT constructs a new DFA (Deterministic Finite-state Automaton)-based model of a faultless PLC control process called Modified DFA, which reflects the signal state vector and timeout function (MDSVTF) model. The first error detection unit 100 detects and isolates defects and abnormalities using the model. The first error detection unit 100 checks whether the observed control process operation matches the modeled control process. The control signal data stored in the PLC memory table (PLC signal log data record) is used for this entire procedure. Each log data is basically an integer vector (or array) containing the (integer) status values of all PLC input/output (I/O) signals. The i-th element of the vector represents the status value of the i-th PLC I/O signal.

2 is a reference diagram illustrating an MDSVTF control process model construction procedure, and FIG. 3 is a reference diagram illustrating an error and anomaly detection and isolation procedure of an MPFDT.

As shown in Fig. 2, initially, an MDSVTF automatic model of the PLC control process is built using the PLC signal log data taken from the defect-free manufacturing system. As shown in Fig. 3, the MDSVTF model is compared with the observed signal log data to detect and isolate defects and anomalies. If there is no difference between the observed and expected behavior of the PLC input and output signals, the system is considered to be error-free and fault-free. Otherwise, it indicates that a defect or anomaly has occurred in the manufacturing system and is reported with detailed diagnostic information.

The first error detection unit 100 includes a modeling module 110 and a first error detection module 120 to perform the above-described functions.

The modeling module 110 models the reference state transition information for a PLC digital input/output signal and a PLC analog input/output signal constituting the PLC input/output signal. To this end, the modeling module 110 performs MDSVTF modeling on reference state transition information including transition timeout time information and signal state vector information corresponding to the PLC digital input/output signal.

4 is a reference diagram illustrating a scenario of a PLC control system for explaining the MDSVTF model building procedure, and FIG. 5 is a signal state versus time chart of the PLC input/output signals (sensor and actuator signals) of the PLC control system shown in FIG. It is a timing diagram exemplifying. In addition, FIG. 6 is a reference diagram illustrating the automatic model of the PLC control system MDSVTF shown in FIG. 4.

5 illustrates a signal state value (0, 1 or other values (100, 150)) at each time. Basically, a specific state of MDSVTF represents a unique vector of integers. The vector contains the integer status values of all PLC input/output signals. As shown in FIG. 6, the following two types of information are attached to each state transition belonging to the MDSVTF model. I) Transition Timeout (TTO) time (each state transition must be completed within the corresponding TTO time); And ii) a signal state vector.

The signal state vector that induces the transition between the two states of the MDSVTF model actually consists of signal state change (SSC) events that occur during that state transition process. The signal name and integer state value may exemplify'_'. For example, the SSC event AWS_100 (meaning that the state value of the AWS signal is updated to 100) is independently responsible for the transition from X1 to X2.

As shown in FIGS. 5 and 6, the modeling module 110 attaches a regular expression in the MPFDT to each MDSVTF model state indicated by a transition going in a plurality of directions. The regular expression equation is used to concisely encode the associated transition execution pattern or motion information. For example, assuming that in state X1 of FIG. 6, all transitions are numbered starting from 0, therefore, transitions X1 → X2 and X1 → X7 are symbolized by numbers 0 and 1, respectively. And they are assumed to be selectively executed by the control system. The regular expression attached to node X1, RE (X1) = (01) *, briefly explains this fact in an efficient way. The regular expression (01) * means that transition 0 (i.e. X1 → X2) is executed first, then transition 1 (i.e., X1 → X7) is executed, and this process repeats until a certain termination condition is reached. Continue. The order or sequence in which transitions from a particular state are executed does not always follow a repeating pattern, and the transition execution operation is characterized by a transition probability value (transition execution/transition occurrence). If the same procedure is performed, a regular expression defining transition time pattern information is attached to each transition having a plurality of TTO times (see FIG. 6 5 ). If there is no regular expression capable of appropriately defining the sequence pattern of the transition time instance, the transition time pattern information may be characterized by a probability value, and accordingly, a lot of control process information may be included in the MDSVTF model.

In addition, the modeling module 110 divides the PLC analog input/output signal into a plurality of segment sections based on a time point at which the PLC analog input/output signal rapidly fluctuates according to a statistical standard, and the reference state transition information including quantization information of the segmented segment sections. Perform MDSVTF modeling for

Most PLC analog input/output signals change their state frequently. The modeling module 110 processes the continuously changing PLC analog input/output signals in the following manner.

In the MDSVTF model, states and transitions are defined based on a change in the state value of a signal. Therefore, if the state value of the PLC analog input/output signal is frequently changed, too many states and transitions may occur in the MDSVT model. This can require a huge amount of state space and memory space. Therefore, the modeling module 110 performs a signal quantization (or data binning) process to greatly reduce the state space complexity of the MDSVTF model.

7 is a reference diagram illustrating a stamping press system among the PLC control systems, and FIG. 8 is a reference diagram illustrating a state versus time chart of a PLC analog input/output signal in the stamping press system shown in FIG. 7. In addition, FIG. 9 is a reference diagram illustrating an MDSVTF model of the stamping press system shown in FIG. 7.

The modeling module 110 performs MDSVTF modeling for the PLC analog input/output signal according to the same procedure as described in FIG. 6. However, the analog sensor signal ADS periodically changes their state values during the time intervals TI1 and TI2 in each signal period. The data points of the signal ADS within these two interval windows need to be quantized and coded to reduce the state space complexity of the model (see Figs. 8 and 9).

As shown in Fig. 8, the time interval TI1 is divided into two discontinuous and adjacent intervals, namely TI11 and TI12. Further, the time interval TI2 is divided into two equal length intervals, TI21 and TI22. Hereinafter, this type of time interval division is referred to as a'marked time interval', and a signal segment within a marked time interval is referred to as a'marked signal segment' (the reader belongs to the marked signal segment. It should be noted that the data points fluctuate significantly). The marked time interval or the boundary of the marked signal segment is determined not only based on the occurrence time of a disruptive SSC event (the analog input/output signal changes its slope sharply afterwards), but also based on the range of state values of the interval. . The modeling module 110 basically automatically determines the boundary of the marked signal segment according to a predefined statistical criterion. All status values of a marked signal segment or PLC analog input/output signal within a marked time interval form a group and are considered as a single SSC event (i.e., the status values within the corresponding status values are quantized to a single status value).

As shown in Fig. 9, the SSC element of the signal state vector is calculated in exactly the same way. However, in the case of PLC analog input/output signals, the status value within the marked signal segment is extended by the status value range label (i.e., the connection of the signal name and the start and end values of the status value range-tokens are separated by'_'). It is represented as a single unit by the signal label. By applying the above signal quantization step, the state space or memory space complexity of the MDSVTF model can be greatly reduced. Here, the SSC event in the form sig_p_q (where sig is the name of the PLC analog input/output signal and p and q are the start and end points of the status value range), the updated status value of the PLC analog input/output signal sig is within the range of [p, q]. Indicates that there should be (see Figs. 8 and 9 8). Transitions in which the signal state vector contains such SSC events have slightly different characteristics. For example, transition x → sig_p_q x'indicates that the MDSVTF of state x acknowledges the SSC event sig_p_q (the updated state value of signal sig is within the valid range) and proceeds to the next state x'. However, until the MDSVTF exits the state x'and moves to the next valid state, the value of the continuous change state of the analog signal sig must be within the range [p, q].

The first error detection module 120 compares the reference state transition information modeled in the modeling module 110 with the detected state transition information to determine whether the PLC control system has an error, and Detect the type of error according to the decision.

For example, when the detected PLC digital input/output signal and the on/off status of the reference state transition information according to MDSVTF modeling are different from each other, the first error detection module 120 causes an error with respect to the detected PLC digital input/output signal. To detect. In addition, the first error detection module 120 detects the detected PLC analog input/output signal as an error when the detected PLC analog input/output signal and the signal state value of the reference state transition information according to MDSVTF modeling do not match. do.

The first error detection module 120 may detect that an error has occurred in the system when a system state transition that is not similar to the state transition of the MDSVTF model is observed. MPFDT can be implemented in the MDSVTF model layout. The first error detection module 120 may detect two types of errors as follows.

i) Transition time error: This type of error occurs when the transition takes longer than the corresponding TTO time.

ii) Transition and Transition Time Probability Errors: This type of error occurs when a significant change is observed within the transition or transition time associated with a particular state or transition.

If PLC input/output signals are found in an error state, an error occurs in the control process. In the case of PLC digital input/output signals, the OFF state is observed instead of the ON state, and vice versa. On the other hand, in the case of PLC analog input/output signals, it means that the range of the observed signal segment does not match the range of the marked signal segment.

Two types of errors can occur in the PLC control process. The first type of error occurs when the control system executes an error or invalid transition. That is, when a transition that does not match the transition of the corresponding MDSVTF model is observed, this type of error occurs more often. The second type of error occurs when the control process is in an infinite loop. In some cases, when a sensor or actuator goes down, the control process gets stuck in an infinite loop without completing subsequent work. Since the first error detection module 120 can observe that the state value of the input/output signal or the system state has not changed for a long time (or there may be slight changes in the case of an analog input/output signal), this type of Errors can be identified.

When an error is detected, the first error detection module 120 identifies all potential error SSC events (error candidate set). For example, in the case of the first type of error, the first error detection module 120 is a set of error candidates generated by MPFDT, and includes all SSC elements of all possible transitions and all SSC elements of error transitions in state X1. Can be detected. That is, the first error detection module 120 may detect an SSC event that is basically expected to occur and an SSC event that actually occurs as a combination of the first type of error candidate. In addition, in the case of a second type of error (the control process corresponds to an infinite loop), the first error detection module 120 selects the actuator SSC element of the last observed transition (instead of the SSC element of the error transition) as an error candidate set. Is detected. This is due to the possibility that some of the actuator actions associated with the last observed state transition did not perform correctly.

Log data obtained from a system controlled by a PLC in the field generally contains a number of data inaccuracies. In short, the data logger may not detect some fast state transitions of the input and output signals. Therefore, in the MDSVTF model, only states belonging to the category of'stable system states' are included. Stable system state refers to a state in which the control process spends (or stops) a long time for the data logger to record complete state information (status values of all input and output signals).

Further, since the state transition always takes a certain amount of time and not in an actual scenario, the maximum value of the observed transition time instances in the TTO time calculation is taken as the actual transition time. For the same reason, due to fluctuations in the status values of the PLC analog input/output signals, the status value range associated with the SSC event of the sig_p_q format can be set wider than the range of all observed marked signal segments.

Using the MDSVTF model described above, errors or anomalies that seriously affect the operation of the system can be reliably detected. However, some soft faults and anomalies that do not significantly affect system operation may not be detected. There can be two reasons the error does not have much effect on the operation of the control process. First, when an error occurs in an independent sensor, it is more of a system architecture design problem and must be addressed in the system architecture design stage. The second is because an actuator error has occurred. Since the actuator signal is an output signal, unlike sensor failure, actuator failure is not reflected in the PLC memory table. Therefore, the control process model may remain unchanged.

For this reason, it is the reason that the second error detection unit 200 needs to detect an error. The second error detection unit 200 divides the power consumption signal according to the control of the PLC control system or the PLC input/output signal according to the control of the PLC control system, and decomposes a plurality of statistical characteristic information and singular values from the divided signals. Whether the PLC control system is abnormal by extracting singular value information by the Value Decomposition (SVD)) algorithm and applying the extracted statistical feature information and the singular value information to a trained autoassociative neural network (AANN). Is detected. The second error detection unit 200 may determine whether there is an error in the system based on whether there is a significant change in the operation of the analog PLC input/output signal or whether there is a significant change in the power consumption profile of the device in the PLC control system.

10 is a reference diagram illustrating an overall error detection procedure performed by the second error detection unit 200. The second error detection unit 200 basically corresponds to a single class classifier based on a self-associative neural network (AANN), and a method capable of detecting important changes in a specific segment of a corresponding time series signal, that is, power consumption or analog input/output signal. It is designed and trained.

The signal segmentation procedure of the second error detection unit 220 generates a total of (N + M) time series databases. N and M databases are each generated by a dividing procedure of a power consumption signal and an analog input/output signal (see FIG. 10). The second error detection unit 220 checks whether the time series signal of the time series database is normal or defective. That is, when a significant part of the time series signal is classified as a defect in a specific time series database, it can be determined that such a failure or anomaly has occurred in the corresponding sensor or actuator. To this end, the second error detection unit 220 may include (N + M) error detection units, one for each time series database. Each of the second error detection units operates independently of each other and all operate correctly in the same manner.

To this end, the second error detection unit 200 includes an input/output signal division module 210, a power signal division module 220, a statistical information detection module 230, a singular value detection module 240, and a second error detection module 250. ).

The input/output signal dividing module 210 divides a PLC analog input/output signal for operation of the PLC control system based on a power cycle. The input/output signal division module 210 transmits the divided PLC analog input/output signals to the statistical information detection module 230 and the singular value detection module 240.

In order to separate the region of interest and maintain the appropriate length of the time series signal, the input/output signal dividing module 210 divides the analog input/output signal into regions (or segments) having similar statistics before being transmitted to the second error detection module.

11 is a reference diagram of an embodiment for explaining a process of dividing an analog input/output signal. Figure 11 (a) illustrates an AOD (Argon Oxygen Decarburization) steelmaking system, and Figure 11 (b) shows that the pressure inside the furnace chamber is measured by an analog pressure sensor called APS, and the signal state of the analog sensor signal APS Here is an example of a time chart.

As shown in (b) of FIG. 11, the input/output signal dividing module 210 divides the entire analog input/output signal into individual energization cycles, and ultimately provides it to the second error detection module 220 for training or classification. The energization cycle starts when the PLC input/output signal is activated or switched on, and the energization cycle ends when deactivated or switched off. Signal segments or regions within the energization cycle window (respectively, the waiting time interval window) are generally referred to as active or inactive signal segments or regions. The input/output signal dividing module 210 determines the analog input/output signal as an active state when it starts to move away from the baseline level, and determines it as an inactive state when it returns to the baseline level.

The input/output signal dividing module 210 divides the analog input/output signal based on the energization cycle based on the following criteria: If a sensor is defective or there is an error in the control process of the actuator, the input/output signal dividing module 210 is Divide when the active area of the input/output signal tends to deviate from the normal than the inactive area. In addition, it is divided when the inactive area of the analog input/output signal is generally too long and needs to be removed from the signal to better capture the period-to-period difference in the signal waveform.

The power signal dividing module 220 divides the power consumption signal provided for the operation of the PLC control system for each device cluster. The power signal dividing module 220 transmits the divided power consumption signal to the statistical information detection module 230 and the singular value detection module 240.

12 is a reference diagram illustrating a process of collecting power usage data.

Referring to Fig. 12, i) a device or machine in a PLC controlled manufacturing system is inherently organized (at the point of electrical connection) in different device clusters. ii) Power usage data (instantaneous power consumption) of each device cluster is continuously recorded (or sampled) by a power meter and inserted into the corresponding power usage log database. Therefore, it is possible to basically have one power usage log database for each device cluster. The power signal dividing module 220 processes power consumption signals (or power use log database) individually and independently. This division procedure for the power consumption signal needs to be repeated according to each device cluster or the power usage log database.

13 is a reference diagram illustrating an operation process of a robot drilling subsystem for a small device cluster, and FIG. 14 is a reference diagram illustrating a power consumption signal profile and a signal division process of the device cluster shown in FIG. 13.

The production system cycle or production cycle begins with the receipt of raw materials and ends with the release of the finished product. As shown in FIG. 14, first, the power signal dividing module 220 determines the power consumption signal of the device cluster corresponding to each production cycle based on a separate set of actuator operations (that is, the actuator set is (Based on being active) it splits it into a number of smaller segments. Each such signal segment is referred to as a task-specific power segment. The task-specific power segment TPS1 basically represents the power usage profile when the actuator is not operating. This is referred to as an inactive segment or region of the power consumption signal.

Thereafter, the power signal dividing module 220 merges or discards the power segment for each task according to the type of actuator that is active (or operated) for a specific time in the next step (see FIG. 14). From a control engineering point of view, a manufacturing system can have two types of actuators. First, as a loosely coupled actuator, it is referred to as a loosely coupled actuator when the operating state of the actuator (eg, normal or faulty) does not have any effect on the operation of the subsequent control process. Second, as a strongly coupled actuator, it is referred to as a strongly coupled actuator when the remaining control process operations cannot be executed normally without successfully completing the corresponding physical operation. MPFDT's MDSVTF model-based error and anomaly search procedures may not detect errors and anomalies related to loosely connected actuators, and these types of errors and exceptions are not effectively handled by the overall control process operation or the MDSVTF model.

Therefore, additional inspection is required for the power segment per operation corresponding to the loosely coupled actuator. As shown in Fig. 14, for each actuator that is loosely coupled in the device cluster, the power segments for each task are merged while the actuator is active, and as a result, the segments are inserted into the corresponding time series database. The remaining task-specific power segments can be ignored or excluded. The signal dividing process may be repeated for the power consumption signal corresponding to each production cycle.

The statistical information detection module 230 subdivides each of the divided power consumption signals or the divided analog input/output signals into three sections, and generates a plurality of statistical equations for each of the divided power consumption signals or analog input/output signals into three sections. Statistical feature information is detected by using.

The statistical information detection module 230 may use 14 statistical equations shown in Table 1 below as statistical equations.

As shown in Table 1, the statistical information detection module 230 is a statistical formula, a mean formula, a standard deviation formula, a root mean square formula, a shape factor formula, Peak Value Formula, Peak-to-Peak Value Formula, Crest Factor Formula, Impulse Factor Formula, Margin or Clearance Factor Formula, Distortion Value ( Each of the statistical feature information is extracted by applying at least one of the skewness) formula, the skewness factor formula, the Kurtosis formula, the Kurtosis factor formula, and the entropy formula. .

FIG. 15 is a flowchart of an embodiment for explaining the procedure of statistical feature extraction, feature selection, and AANN learning performed by the second error detection unit 200 as a whole. In addition, FIGS. 16A to 16D are partial flowcharts for explaining the flowchart illustrated in FIG. 15 in detail.

As the output of the feature extraction process, it basically contains a feature database or a P × 17 matrix. Where P is the number of time series signals in the time series database. The statistical information detection module 230 extracts ten statistical feature values from each time series signal, and the singular value detection module 240 detects seven singular values as described later.

16A is a partial step of a flowchart illustrating an operation process of the statistical information detection module 230.

Referring to FIG. 16A, the statistical information detection module 230 extracts features of a signal using a plurality of statistical equations described in Table 1 above. The statistical information detection module 230 may select a formula capable of clearly distinguishing whether or not there is an error from among the used statistical formulas. The statistical information detection module 230 subdivides an analog input/output signal or a power consumption signal into three signals. In general, it is possible to obtain the most distinct features by dividing it into three. The statistical information detection module 230 selects a statistical equation for each of the three subdivided signals. First, the statistical information detection module 230 calculates a value by substituting 14 statistical equations into each signal. After that, the statistical information detection module 230 selects the top three equations that can better distinguish the abnormal signal from among the 14 statistical equations substituted for each signal. Since 3 values are selected from each signal, a total of 9 values may be selected. As a criterion for selecting the top three equations, the statistical information detection module 230 selects an equation with little fluctuation in the result value for all signals, or when comparing a signal in a normal state with a signal when an error occurs, Either an equation that provides the difference between the result values is selected, or values in the same cluster are clustered, and as the distance between different clusters increases, an equation with better characteristics is selected.

The singular value detection module 240 extracts singular value information using a singular value decomposition (SVD) algorithm from the subdivided signal.

16B is a partial step of a flowchart illustrating an operation process of the singular value detection module 240.

Referring to FIG. 16B, the singular value detection module 240 searches for a singular value that is difficult to extract in a statistical manner using the SVD algorithm. For example, the singular value detection module 240 converts an analog input/output signal or a power consumption signal into a matrix having a size of 11 rows. After that, the singular value detection module 240 extracts 11 results by substituting each row of the transformed matrix into a SVD (Singular Value Decomposition) algorithm. The singular value detection module 240 selects the top 7 results from among 11 results. The selection is based on the result that can best distinguish between normal and abnormal data.

The second error detection module 250 detects whether the PLC control system is abnormal by performing a homogeneity mapping process using the self-associative neural network (AANN) algorithm on the extracted statistical feature information and the singular value information.

16C and 16D are some steps of a flowchart illustrating an operation process of the second error detection module 250.

As shown in FIG. 16C, the second error detection module 250 stores 17 features including 10 statistical feature information including one time series length value and 7 singular value information in a database. At this time, the second error detection module 250 stores one length vector of the signal together.

Thereafter, as shown in FIG. 16D, the second error detection module 250 learns features stored in the database through the self-associative neural network (AANN) algorithm. The second error detection module 250 may store the learned result in a separate database, and may detect an error of the PLC control system by comparing the information registered and observed information through the learned neural network.

The second error detection module 250 includes two databases, that is, a reference time series database and a query time series database. The reference time series database contains a large number of normal time series signals. The second error detection module 250 learns by using a sample of the reference time series database. On the other hand, the query time series database contains a small number of samples extracted from the observed analog input/output signals or power consumption signals. The second error detection module 250 determines whether the time series signal in the time series database is normal or error. To this end, the second error detection module 250 includes an AANN-based signal deviation detection algorithm for each time series database.

The second error detection module 250 needs to be trained using the feature vector of the reference feature database. When the second error detection module 250 is fully trained, it can be applied to classify a time series signal in a query time series database. The second error detection module 250 uses AANN as a time series classifier. AANN is basically a kind of self-directed feed forward neural network designed to perform homogeneity mapping. If the network is fully trained using feature vectors extracted from samples of a particular class, the reconstruction error for input feature vectors extracted from samples of the same class is expected to be significantly lower than that of samples of other classes. Therefore, the distance between the input vector and the output vector of AANN (commonly referred to as NOVELTY INDEX) can be used as an index to characterize the state of the input sample. Also, by determining an appropriate threshold, the test or query samples can be classified into normal or error groups.

The following is an example of a reference time series database and a query time range database.

Case 1: Active signal segment of ADS of analog sensor signal shown in FIG. 8

The reference time series database may include 200 normal signal samples. Signal samples are generated by simulating the stamping press system of FIG. 7. The simulator is programmed in such a way that each energization cycle of the analog sensor signal ADS is completed within 60 to 61 seconds.

The corresponding query time series database may include 90 signal samples. The first 60 samples are normal signal samples and the remaining 30 samples are defective signal samples. Standard signal samples are generated according to the same procedure as described in the reference time series database. To generate a sample of the error signal, the simulator is reprogrammed in such a way that it takes 61.5 to 62.5 seconds for each energization cycle of the analog sensor signal ADS to complete. In practice this can be done by randomly slowing down or increasing the speed of the stamping press actuator. The speed up or deceleration mode of the stamping press system can be started and completed at any time within the corresponding energization cycle time.

Case 2: Active signal segment of ADS of analog sensor signal shown in FIG. 11

The reference time series database may include 200 normal signal samples. Signal samples are generated by simulating the AOD stainless steel system of FIG. 11. The simulator is programmed in such a way that the pressure inside the furnace drops from 0.7 (± 0.02) atmosphere (atm) to 0.5 (± 0.02) atmosphere in 120 (± 2) second intervals.

The corresponding query time series database may include 90 signal samples. The first 60 samples are normal signal samples and the remaining 30 samples are defective signal samples. Normal signal samples are generated according to the same procedure as described in the reference time series database. To generate a sample of the error signal, the simulator is reprogrammed in such a way that the pressure inside the furnace drops from 0.76 (± 0.03) air pressure to 0.44 (± 0.03) air pressure in 121 (± 3) second intervals.

Case 3: Power consumption signal segment of the robotic drilling system shown in FIG. 14

The reference time series database contains 200 normal signal samples. The power consumption signal is collected by running the spindle motor for 3 minutes, and this process can be repeated 40 times to obtain 40 power consumption samples. Another 160 signal samples are generated by adding small random noise to the 40 signal samples.

The corresponding query time series database contains 40 signal samples. The first 30 samples are normal signal samples and the remaining 10 samples are defective signal samples. Normal signal samples are generated according to the same procedure as described in the reference time series database.

17A to 17C are reference diagrams illustrating classification performance of the second error detection unit 200 according to the present invention. 17A shows the classification performance of the second error detection unit 200 for the query time series database of Case 1, and FIG. 17B shows the classification performance of the second error detection unit 200 for the query time series database of Case 2, 17C is a reference diagram showing the classification performance of the second error detection unit 200 for the query time series database in Case 3;

17A to 17C, the second error detection unit 200 can accurately classify almost all samples. As can be seen from FIG. 17A, the second error detection unit 200 can accurately classify most of the samples except for two. In the case of case 1, the second error detection unit 200 cannot correctly classify several samples because the defective sample used in this case study is not significantly different from the normal sample. However, in Cases 2 and 3, realistic training and test samples were used, and it can be seen that the second error detection unit 200 accurately classifies all samples.

The present invention can be implemented by a computer-readable recording medium by implementing a software program. For example, the recording medium may be a hard disk, flash memory, RAM, ROM, etc. as a built-in type of each playback device, or an optical disk such as a CD-R or CD-RW, a compact flash card, a smart media, a memory stick, or a multimedia card as an external type. have.

Although the embodiments of the present invention have been described above, the embodiments disclosed in the specification of the present invention are not intended to limit the present invention. The scope of the present invention should be interpreted by the following claims, and all technologies within the scope equivalent thereto should be interpreted as being included in the scope of the present invention.

100: first error detection unit
110: modeling module
120: first error detection module
200: second error detection unit
210: I/O signal division module
220: power signal division module
230: statistical information detection module
240: singular value detection module
250: second error detection module

Claims (10)

The state values at the same time for each of a plurality of device operations according to the PLC input/output signal of the PLC (Programmable Logic Controller) control system are detected as state transition information for the PLC control system, and the detected state transition information and the A first error detection unit that compares reference state transition information during normal operation of the PLC control system to detect whether or not there is an error in the PLC control system; And
The power consumption signal according to the control of the PLC control system or the PLC input/output signal according to the control of the PLC control system are divided, a plurality of statistical feature information is extracted from the divided signal, and the extracted statistical feature information is trained. Including a second error detection unit for detecting the abnormality of the PLC control system by applying to an associative neural network (ANNN: Autoassociative Neural Network),
The first error detection unit,
A modeling module for modeling the reference state transition information for a PLC digital input/output signal and a PLC analog input/output signal constituting the PLC input/output signal; And
The first error for determining whether or not the PLC control system has an error by comparing the reference state transition information modeled in the modeling module with the detected state transition information, and detecting an error type according to the error determination for the PLC control system. Including a detection module,
The modeling module,
An error detection system according to PLC control, characterized in that performing MDSVTF modeling on the reference state transition information including transition timeout time information and signal state vector information corresponding to the PLC digital input/output signal. delete delete The method according to claim 1,
The modeling module,
Dividing the PLC analog input/output signal into a plurality of segment segments based on a time point at which the PLC analog input/output signal rapidly fluctuates according to a statistical standard, and performing MDSVTF modeling on the reference state transition information including quantization information of the segmented segment segments. Error detection system according to PLC control, characterized by. The method according to claim 1,
The first error detection module,
Error detection system according to PLC control, characterized in that when the detected PLC digital input/output signal and the on/off status of the reference state transition information according to MDSVTF modeling are different, the detected PLC digital input/output signal is detected as an error. . The method according to claim 1,
The first error detection module,
Error detection according to PLC control, characterized in that when the detected PLC analog input/output signal and the signal status value of the reference state transition information according to MDSVTF modeling do not match, the detected PLC analog input/output signal is detected as an error. system. The method according to claim 1,
The second error detection unit,
An input/output signal dividing module for dividing a PLC analog input/output signal for operation of the PLC control system based on an energization cycle; And
And a power signal dividing module for dividing the power consumption signal provided for the operation of the PLC control system for each device cluster. The method of claim 7,
The second error detection unit,
The divided power consumption signal or each of the divided analog input/output signals is subdivided into three sections, and the average value formula, standard deviation formula, and square mean for each of the power consumption signals or the analog input/output signals subdivided into three sections Any one or more of formula, shape factor formula, peak value formula, peak-to-peak value formula, crest factor formula, impulse factor formula, redundant factor formula, distortion value formula, distortion factor formula, sharpness formula, sharpness factor formula, and entropy formula And a statistical information detection module for extracting each of the statistical feature information by applying a statistical formula. The method of claim 7,
The second error detection unit,
Error according to PLC control, characterized in that it comprises a singular value detection module for extracting singular value information by a singular value decomposition (SVD) algorithm from the divided power consumption signal or the divided analog input/output signal. Detection system. The method of claim 9,
The second error detection unit,
And a second error detection module configured to detect an abnormality in the PLC control system by performing a homogeneity mapping process using the self-associative neural network (AANN) algorithm on the extracted statistical feature information and the singular value information. Error detection system according to PLC control.

Images