JP6427936B2 - Manufacturing process operation support method, apparatus and program - Google Patents

Description

  The present invention relates to an operation support method, apparatus, and program suitable for supporting operation of a manufacturing process in which a product flows, such as a continuous casting process.

In order to realize a stable operation of continuous casting, it is necessary for an operator to always judge and select an appropriate operation. For that purpose, it is important to accurately predict the state of equipment and products.
For the above prediction, operators, particularly experts, have constructed a cognitive model of the continuous casting process by integrating piecewise information about the state of steel in the continuous casting process through causal relationships. And since operators have knowledge of the cognitive model of the continuous casting process in their heads, they have avoided trouble by estimating the state of the steel, predicting the future state, and performing appropriate operations.
On the other hand, in recent years, productivity decline due to a decrease in skilled workers, operation trouble occurrence, and the like are increasing, and it is demanded to support appropriate operations even if they are not skilled workers.

Focusing on the cognitive model in the expert's mind, it was also in the mind of the expert so far, taking into account the complex mutual interference of equipment and product states as described above. There is a need to make the cognitive model of the continuous casting process objective, that is, to model causal relationships in the state transition of steel so that it can provide guidance for avoiding operational troubles and taking actions when operational troubles occur.
Therefore, the present applicant proposes an operation support apparatus, method, and program suitable for supporting the operation of a manufacturing process in which a continuous product continuously flows as in the continuous casting process in Patent Document 1. Yes. Specifically, the continuous casting process is divided into multiple stages in the product flow direction (from the upstream side, tundish stage, nozzle stage, mold stage, vertical area stage), and the state transition model of product attributes for each stage And a state transition model of equipment, and discrete modeling is performed by defining the influence relationship in each stage. Then, the reachable tree creation unit sets an initial state for each stage using the discrete model, creates a reachable tree that is a state transition diagram of each stage indicating a state that can be reached from the initial state, The total reachable tree creating unit creates a reachable tree for the entire continuous process using the reachable trees of all stages created by the reachable tree creating unit. Furthermore, the risk matrix creation unit creates a risk matrix that represents the relationship between the occurrence probability and the risk using the reachable tree of the entire continuous casting process created by the overall reachable tree creation unit.

JP 2012-146269 A

In Patent Document 1, a reachable tree indicating a reachable state from the initial state is created. However, when the initial state is unknown, there is a problem that the reachable tree cannot be created.
In addition, a risk matrix that represents the relationship between the occurrence probability and the risk level is created. However, since the probability of state transition is not taken into account, there is a problem that it is difficult to grasp the rate of occurrence of trouble that tends to fall.

  The present invention has been made in view of the above points, and it is possible to create a state transition diagram by appropriately setting an initial state and to evaluate the likelihood of falling into trouble in consideration of the probability of state transition. The purpose is to do so.

According to the manufacturing process operation support method of the present invention, a manufacturing process in which a product flows is divided into a plurality of stages in the flow direction of the product, and a state transition model of an attribute of the product and a state transition model of an equipment are provided for each stage. And supporting the operation of the manufacturing process based on the discrete model created by defining the influence relationship between the state of the product attribute and the state of the equipment in each stage. In this operation support method, the means for setting the initial state is observable acquired in the current operation using the state transition probability obtained from the state transition frequency of the unobservable state and the state transition frequency of the observable state. a step of setting an initial state by Bayesian estimation in a manner consistent to the operational numerical data, means for determining the existence probability, based on the discrete model, the initial state of the set Creating a full state transition diagram, and using the state transition frequency of the unobservable state and the state transition probability determined from the state transition frequency of the observable state, to obtain the existence probability of a predetermined state. Features.
Another feature of the operation support method of the manufacturing process according to the present invention is that the manufacturing process is a manufacturing process in which a continuous product continuously flows.
Another feature of the manufacturing process operation support method of the present invention is that the state transition frequency of the unobservable state is given in advance, and the observable state has one factor in the state transition of the observable states. The state transition frequency of the state is obtained from the actual value of the observable operation numerical data, and the state transition frequency of the observable state having a plurality of factors in the state transition among the observable states is the actual value of the observable operation numerical data. Based on the initial existence probability of the observable state, the initial existence probability of the unobservable state given in advance, and the state transition frequency of the unobservable state, obtain the probability of state transition from the state transition frequency of the unobservable state, Each time there is a state transition of the observable state, it is in a point that is obtained by establishing simultaneous equations.
In addition, another feature of the manufacturing process operation support method of the present invention is that in the step of setting the initial state,
P (A | B) = P (B | A) P (A) / P (B)
In Bayesian estimation represented by
The posterior probability P (A | B) is a probability that an unobservable state of interest exists at the predetermined time under the condition that operation numerical data from a predetermined time to the present is observed, and P ( B | A) P (A) is a probability that the observed unobservable state exists at the predetermined time and the operation numerical data from the predetermined time to the present is observed, and P (B ) Is a probability that the operation numerical data from the predetermined time to the present is observed, and when the operation numerical data from the predetermined time to the present is given as evidence, pay attention to a certain unobservable state. Thus, the existence probability of the unobservable state of interest at the predetermined time is obtained as the posterior probability P (A | B).
Further, another feature of the manufacturing process operation support method of the present invention is that the discrete model is represented by a directed graph in which state transition is caused by transition firing, and firing of a transition corresponding to the probability of the state transition is performed. Assuming that the probability is an exponential distribution of the firing frequency of the transition corresponding to the state transition frequency, the probability that the transition will fire during a predetermined time is represented by a cumulative distribution function.
The manufacturing process operation support device according to the present invention divides a manufacturing process in which a product flows into a plurality of stages in the flow direction of the product, and a state transition model of an attribute of the product and a state transition model of an equipment for each stage. And supporting the operation of the manufacturing process based on the discrete model created by defining the influence relationship between the state of the product attribute and the state of the equipment in each stage. Using the state transition frequency of the unobservable state and the state transition frequency obtained from the state transition frequency of the observable state, so as not to contradict the observable operation numerical data acquired in the current operation. Based on the means for setting an initial state by Bayesian estimation and a state transition diagram starting from the set initial state based on the discrete model, the state transition frequency of the unobservable state Using the probability of the state transition obtained from the state transition frequency of the observable state, characterized by comprising a means for determining the existence probability of a given state.
The program of the present invention divides a manufacturing process through which a product flows into a plurality of stages in the flow direction of the product, creates a state transition model of the attribute of the product and a state transition model of equipment for each stage, A program for supporting the operation of the manufacturing process based on a discrete model created by defining an influence relationship between the state of the attribute of the product and the state of the equipment in each stage. Using the state transition frequency obtained from the state transition frequency of the unobservable state and the state transition frequency of the observable state, the initial state is set by Bayesian estimation so as not to contradict the observable operation numerical data acquired in the current operation. And a state transition diagram starting from the set initial state based on the discrete model, the state transition frequency of the unobservable state and the observable state Using the probability state obtained from the state transition frequency transition, to execute a process of determining the existence probability of a given state the computer.

  According to the present invention, the state transition diagram can be created by setting the initial state by estimating the existence of the unobservable state so as not to contradict the observable operation numerical data acquired in the current operation. The probability of falling into trouble can be evaluated in consideration of the probability of.

It is a figure which shows the structure of the operation assistance apparatus of the continuous casting process which concerns on embodiment. It is a figure which shows the example of the continuous casting model which carried out discrete modeling of the continuous casting process. It is a figure for demonstrating a discrete state record. It is a figure which shows the example of the state transition obtained from the track record value of observable operation numerical data. It is a figure which shows the example of a reachable tree. It is a characteristic view which shows probability density distribution and a cumulative distribution function. It is a figure which shows the state transition which is marking 1 in the time t = 0, and becomes marking 2 in the time t = T. It is a figure which shows the state transition which is marking 1 in time t = 0, passes marking 2, and becomes marking 3 in time t = T. It is a figure which shows the example of a reachable tree. It is a figure which shows the example of the reachable tree produced in a dangerous state probability calculation part. It is a figure which shows an example of the guidance information which the operation support apparatus of a continuous casting process shows to a user. It is a figure which shows the outline | summary of a refining process. It is a figure explaining the attribute of the product of each stage.

Preferred embodiments of the present invention will be described below with reference to the accompanying drawings.
When applying the present invention, using the method of Patent Document 1, the continuous casting process is divided into a plurality of stages (tundish stage, nozzle stage, mold stage, vertical region stage from the upstream side) in the flow direction of the product, A state transition model of product attributes and a state transition model of equipment are created for each stage, and discrete modeling is performed by defining the influence relationship in each stage. A discrete model (continuous casting model) obtained by discrete modeling of the continuous casting process is represented by, for example, a Petri net that is a directed graph.

As shown in FIG. 2, in a continuous casting model in which the continuous casting process is discretely modeled, the temperature at the tundish stage (hereinafter referred to as “tundish steel temperature”) (usually low), the temperature at the nozzle stage (hereinafter referred to as “tundish steel temperature”). (Referred to as “nozzle steel temperature”) (usually low), nozzle at the nozzle stage (hereinafter referred to as “nozzle clogging”) (usually narrow), viscosity at the TD stage (hereinafter referred to as “steel viscosity”) ) (Usually high) is defined.
In this continuous casting model, the temperature of the tundish steel has one factor in the state transition, and the state transitions from place (normal) to place (low) when one transition is ignited. The same applies to the nozzle steel temperature and the steel viscosity. On the other hand, nozzle clogging has a plurality of factors in the state transition, and the state transitions from the place (normal) to the place (narrow) when any of the transitions T ₁ , T ₂ , T ₃ ignites. Transition T ₂ ignites when the state of steel viscosity changes from normal to high. Transition T ₃ ignites when the nozzle steel temperature transitions from normal to low. The transition T ₁ is ignited by other factors.

In FIG. 1, the structure of the operation support apparatus of the continuous casting process which concerns on embodiment is shown.
Reference numeral 1 denotes a database that stores actual values of observable operation numerical data. In the continuous casting process, there are an observable state in which the state can be observed by the operation numerical data, and an unobservable state in which the state cannot be observed by the operation numerical data. For example, tundish steel temperature and nozzle clogging are observable. On the other hand, nozzle steel temperature and steel viscosity are unobservable. The database 1 stores actual values of observable operation numerical data (for example, tundish steel temperature [° C.]).

  Reference numeral 2 denotes a discrete numerical actual value data creation unit that creates a discrete state record 101 from the actual values of the observable operational numerical data stored in the database 1. First, from the actual values of the observable operation numerical data stored in the database 1, as shown in FIG. 3A, discrete states representing the states of the observable states A, B, C,. Create a record. For example, A is the tundish steel temperature (normal), B is the tundish steel temperature (low), C is the nozzle clogged (normal), and so on. The actual value of the observable operation numerical data is a continuous value (for example, a tundish steel temperature [° C.]), but a threshold value for dividing it (for example, a threshold value for dividing the tundish steel temperature (usually low)) is given in advance. . Here, referring to the discrete state record of FIG. 3A, there is no state transition at times t0 and t1, and a state transition occurs at time t2. Therefore, as shown in FIG. 3B, those without state transition are collectively referred to as a discrete state record 101.

  Reference numeral 3 denotes an observable firing frequency calculation unit that calculates an observable firing frequency 104. The firing frequency is the number of times (expected value) that a transition fires per unit time.

  The firing frequency 104 of the tundish steel temperature in the observable state is obtained from the discrete state record 101. As shown in FIG. 2, the tundish steel temperature has one factor in the state transition, and the state transitions from a place (normal) to a place (low) when one transition is ignited. If there is only one transition, it is possible to know in which period it is normal and in which period it is low from the actual value, so that it is possible to know the transition firing frequency. FIG. 2 illustrates the transition from nozzle clogging (normal) to nozzle clogging (narrow), and only the transitions related to the transition are described.

The firing frequency 104 (T ₁ firing frequency, T ₂ firing frequency, T ₃ firing frequency) of the nozzle clogging in the observable state is the actual value of the observable operation numerical data for a certain period in the discrete numerical actual value data creation unit 2 The firing probability is obtained from the firing frequency 103 in the unobservable state based on the discrete state record 101 created from the above, the initial existence probability 102 in the unobservable state, the firing frequency 103 in the unobservable state, and the initial existence probability 106 in the unobservable state. As described below, each time there is a state transition of an observable state, a simultaneous equation is established. The initial existence probability 102 of the unobservable state is given in advance, for example, an empirical probability of what the unobservable state is at the start of casting based on past operational knowledge or the like. Further, the firing frequency 103 in the unobservable state is also given in advance based on past operational knowledge and the like.

  According to the discrete state record 101, as shown in FIG. 4, at the time of casting start (0:00 [minute]), the temperature of the tundish steel (low) and nozzle clogging (normal) were detected. At 0:03 [min], there was no state transition of tundish steel temperature (low) and nozzle clogging (normal), but the steel drawing speed, which is a factor affecting all stages, was changed. At 0:08 [min], the state transitions to tundish steel temperature (low) and nozzle clogging (narrow).

  The initial observation probability 102 of the unobservable state is that the probability of the nozzle steel temperature (normal) and the steel viscosity (normal) is 65 [%] when the tundish steel temperature (low) and nozzle clogging (normal) are present. The probability of nozzle steel temperature (normal) and steel viscosity (high) is 5%, the probability of nozzle steel temperature (low) and steel viscosity (normal) is 25%, nozzle steel temperature (low) and The probability of steel viscosity (high) is given as 5 [%].

FIG. 5 is an example of a reachable tree that is a state transition diagram showing a state that is transitionable (reachable) from the initial state. A reachable tree is created based on a continuous casting model, and a “state” (hereinafter referred to as a marking) marked with an ellipse is marked with a “transition (also called a sequence. It corresponds to the contents of an operation). ”Is connected.
Markings 1 to 4 represent initial states, and P _{1 , 1 to} P _{1, and 4} are initial existence probabilities of the observable state and the unobservable state, respectively. In the example of FIG. 4, for example, the marking 1 represents a state of “tundish steel temperature (low), nozzle steel temperature (normal), nozzle clogging (normal), steel viscosity (normal)”, and P _{1 and 1} are “65 [%]”. Marking 2 indicates the state of “tundish steel temperature (low), nozzle steel temperature (normal), nozzle clogging (normal), steel viscosity (high)”, and P _{1 and 2} are “5 [%]”. Become. Marking 3 indicates the state of “tundish steel temperature (low), nozzle steel temperature (low), nozzle clogging (normal), steel viscosity (normal)”, and P _{1 and 3} are “25 [%]”. Become. In addition, the marking 4 represents the state of “tundish steel temperature (low), nozzle steel temperature (low), nozzle clogging (normal), steel viscosity (high)”, and P _{1 and 4} are “5 [%]”. Become.
The markings 5 to 8 represent states that can be transitioned from the markings 1 to 4, and P _{2, 1 to} P _{2 and 4} are the existence probabilities of the observable state and the unobservable state, respectively.
FIG. 5 illustrates an example of the transition between markings and does not show all possible transitions exhaustively.

A method of obtaining the nozzle clogging firing frequency 104 (T ₁ firing frequency, T ₂ firing frequency, T ₃ firing frequency) will be described. Using the least square method, the firing frequency 104 (T ₁ firing frequency, T ₂ firing frequency, T ₃ firing frequency) is obtained. Here, for the sake of simplicity of explanation, the calculation period is from 0:00 [minutes] to 0:08 [minutes] in the example of FIG.
The expected value of the number of firings between 0:00 [minutes] and 0:03 [minutes] is
3 [minutes] x 0.65 x T ₁ firing frequency (1)
3 [minutes] × 0.05 × (T ₁ firing frequency + T ₂ firing frequency) (2)
3 [minutes] x 0.25 x (T ₁ firing frequency + T ₃ firing frequency) (3)
3 [minutes] × 0.05 × (T ₁ firing frequency + T ₂ firing frequency + T ₃ firing frequency) (4)
It becomes.

Next, the firing probability is obtained from the firing frequency 103 of the unobservable state (how to determine the firing probability will be described later), and the existence probability of each unobservable state at 0:03 [minutes] is obtained. The existence probability of the unobservable state at 0:03 [minutes] is the existence probability of the unobservable state at 0:00 [minutes] capable of state transition to the unobservable state (initial existence probability 102 of the unobservable state). Can be obtained by multiplying by the firing probability. When there are a plurality of unobservable states at 0:00 [minutes] in which state transition to the unobservable state is possible, the existence probabilities obtained from the respective states may be added. Referring to FIG. 5, it is assumed that the markings 1 to 4 are in a state of 0:00 [minutes], the markings 5 to 8 are in a state of 0:03 [minutes], and the existence probability of the marking 8 is obtained. The markings that can transition to the marking 8 are the markings 3 and 4 (existence probabilities P _{1, 3} , P _{1, 4} ). Each solid arrow connecting the markings 3 and 4 and the marking 8 is given an ignition probability p _{3 → 8} , p _{4 → 8} . In this case, the existence probabilities P _{2 and 4} of the marking 8 are P _{1, 3} × p _{3 → 8} + P _{1 and 4} × p _{4 → 8} .
As a result, as shown in FIG. 4, when the tundish steel temperature (low) and the nozzle clogging (normal), the nozzle steel temperature (normal) and the steel viscosity (normal) have a probability of 55 [%], the nozzle Probability of steel temperature (normal) and steel viscosity (high) is 5 [%], probability of nozzle steel temperature (low) and steel viscosity (normal) is 35 [%], nozzle steel temperature (low) and steel viscosity The probability of being (high) was 5 [%]. Note that the firing frequency 103 in the unobservable state is manually provided, and therefore the total probability of existence of the unobservable state determined based on the firing frequency 103 is not always 100%. Therefore, normalization is performed so that the total is 100 [%].
Therefore, the expected value of the number of firings between 0:03 [min] and 0:08 [min] is
5 [minutes] x 0.55 x T ₁ firing frequency (5)
5 [minutes] × 0.05 × (T ₁ firing frequency + T ₂ firing frequency) (6)
5 [minutes] x 0.35 x (T ₁ firing frequency + T ₃ firing frequency) (7)
5 [minutes] × 0.05 × (T ₁ firing frequency + T ₂ firing frequency + T ₃ firing frequency) (8)
It becomes.

The formulas (1) to (8) thus established are summarized. Specifically, both sides of Expression (1) and Expression (5) are added to obtain Expression (9). The other three sets of equations are added in the same manner to obtain equations (10) to (12). That is, the expected value of the number of firings is obtained for all sets of simultaneous firing transitions (only transition T ₁ , transition T ₁ + T ₂ , transition T ₁ + T ₃ , transition T ₁ + T ₂ + T ₃ ).
(3 × 0.65 + 5 × 0.55) × T ₁ firing frequency (9)
(3 × 0.05 + 5 × 0.05) × (T ₁ firing frequency + T ₂ firing frequency) (10)
(3 × 0.25 + 5 × 0.35) × (T ₁ firing frequency + T ₃ firing frequency) (11)
(3 × 0.05 + 5 × 0.05) × (T ₁ firing frequency + T ₂ firing frequency + T ₃ firing frequency) (12)
Using the number of firings of the simultaneous firing transitions observed in operations (9) to (12) (transition T ₁ only, transition T ₁ + T ₂ , transition T ₁ + T ₃ , transition T ₁ + T ₂ + T ₃ ) only the set (transition T ₁ of the simultaneous firing transitions to the square sum of the difference between the expected value of the ignition times and ignition times observed by operation is minimized, the transition T ₁ + T _2, the transition T ₁ + T _3, the transition T ₁ + T determining the expected value of the firing times of ₂ + T _3). Thereafter, T ₁ firing frequency, T ₂ firing frequency, and T ₃ firing frequency are solved as simultaneous equations. In the case of this example, since the right side of all the expressions is 0 (the number of firings), the T ₁ firing frequency, the T ₂ firing frequency, and the T ₃ firing frequency are all 0.
As described above, the expected value of the number of firings is obtained every time there is a state transition in the observable state, and by solving the expected value using the least square method, the nozzle clogging firing frequency 104 (T ₁ firing frequency, T ₂ firing) frequency, determine the T ₃ firing rate).

Here, how to determine the transition firing probability will be described. The firing probability of the transition can be obtained from the firing frequency 103 in the unobservable state and the firing frequency 104 in the observable state.
The transition firing probability is assumed to be an exponential distribution of the transition firing frequency λ. That is, the transition firing probability is a distribution given by the probability density distribution of the following equation (13) with respect to the transition firing frequency λ (see FIG. 6A), and the transition between time 0 and time t The probability (cumulative distribution function) that fires is expressed by the following equation (14) (see FIG. 6B). Here, as the firing frequency λ, the value of the observable firing frequency 104 obtained from the actual value is used. When there are a plurality of actual values of the firing frequency for the same transition, an average value of these firing frequency values may be used.

As shown in FIG. 7, the probability of marking 1 at time t = 0 and marking 2 at time t = T is obtained. In FIG. 7, T _1-1 ~T _1-n1 are those cited all ignitable transitions in the marking 1. λ _{1-1 to} λ _1-n1 are average firing frequencies of transitions that can be ignited in the marking 1. T _{2-1 to} T _2-n2 list all the transitions that can be ignited in the marking 2. λ _{2-1 to} λ _2-n2 are average firing frequencies of transitions that can be ignited in the marking 2.
The timing of the transition fires and t = t _f, transitions ignited at time t = t _f will be described a case is T _1-1.

Marking 1 at time t = 0, the probability of transition T _1-1 transitions to the marking 2 firing between t _f ~t _{f +} Δ, as shown in the following equation (15), "t = t _f The transition T _{1-2 to} T _1-n1 does not fire until the transition T _1-1 fires between t _{f and} t _{f + Δt} , and the transition T ₂ from t = t _f to t = T _{−1 to} T _2−n2 can be obtained by multiplying by “the probability that no firing occurs”. Therefore, when marking 1 is performed at time t = 0, the probability of marking 2 at time t = T is expressed by the following equation (16). Note that sumofλ ₂ means the sum of λ _{2-1 to} λ _2-n2 .

Next, as shown in FIG. 8, the probability of marking 1 at time t = 0, transition 2 ignited, passing through marking 2, and marking 3 at time t = T is obtained. Considering the same as the case of FIG. 7, if the timings when the transitions are fired are t ₁ and t ₂ and the minute times Δ ₁ and Δ ₂ are defined, the probability that the state transitions at the timings t ₁ and t ₂ is 17), the probability to be obtained is given by the following equation (18). In addition, it is expandable also when the frequency | count of a state transition is 3 times or more.

  4 is a discrete numerical current operation data creation unit, which creates a discrete state record 105 from the operation values of the observable operation numerical data acquired in the current operation. The method of creating the discrete state record is the same as that of the discrete numerical result value data creation unit 2, and the description thereof is omitted here.

  Reference numeral 5 denotes an initial state setting unit, which is a discrete state record 105 created by the discrete numerical current operation data creation unit 4, an observable state firing frequency 104 computed by the observable state firing frequency computation unit 3, and an unobservable state initial stage. Based on the existence probability 102, the firing frequency 103 in the unobservable state, and the initial existence probability 106 in the observable state, using the firing probability obtained from the firing frequency 103 in the unobservable state and the firing frequency 104 in the observable state, Bayes The initial state is set by estimation. The initial existence probability 106 of the observable state can be statistically obtained from the actual value of the operation numerical data.

Bayesian estimation is expressed by the following equation (19).
P (A | B) = P (B | A) P (A) / P (B) (19)
P (A) is the probability (prior probability) of event A before event B occurs, and P (A | B) is the probability (event probability) of event A after event B occurs. Using Bayesian estimation, the posterior probability P (A | B) reflects that if a certain result (data) about event B is obtained, it is reflected by multiplication of likelihood P (B | A). Probabilities are updated from prior probabilities to posterior probabilities.
In the description according to the present embodiment, when operation numerical data (observable data) from a predetermined time T (for example, at the start of casting) to the present is given as evidence (corresponding to event B occurring) Further, paying attention to a certain unobservable state (corresponding to event A), the existence probability of the unobservable state focused on at time T is obtained as a posteriori probability P (A | B). In other words, the posterior probability P (A | B) is a probability that an unobservable state of interest exists at time T under the condition that operation numerical data (observable data) from time T to the present is observed. is there. P (B | A) P (A) is the probability that the unobservable state of interest exists at time T and the operation numerical data (observable data) from time T to the present is observed. Become. P (B) is a probability that operation numerical data (observable data) from time T to the present will be observed.

As shown in FIGS. 9A and 9B, all the initial states that can exist at the time T are set based on the initial existence probability 102 of the unobservable state and the initial existence probability 106 of the observable state. Based on the casting model, a reachable tree is created at regular time intervals from the initial state to the present. In the example of FIGS. 9A and 9B, it is assumed that there is only one observable state and only one unobservable state, and () of each marking represents (observable, unobservable) and 1 exists. Means that 0 does not exist. Based on the initial existence probability 102 of the unobservable state and the initial existence probability 106 of the observable state, the existence probability of the initial state is P _{1, 1} , P _{1, 2} , P _{1, 3} , P _{1, 4 in} order from the top. As given.
Further, it is assumed that the observable state observed from the discrete state record 105 changes from 0 (time T) → 1 (time T1) → 0 (time T2 (current)).

A method for obtaining the denominator P (B) of Expression (19) will be described. As shown in FIG. 9A, considering state transitions consistent with the observable state observed from the discrete state record 105, from the markings (0, 0), (0, 1) indicated by solid lines in the initial state, This is a path through which the state transitions from marking (1, 0), (1, 1) at time T1 to marking (0, 0), (0, 1) at time T2. Each solid arrow, firing probability _{p 1, 1 → 2, 3} , p 1, 1 → 2, 4, p 1, 2 → 2, 3, p 1, 2 → 2, 4 are applied. The method for obtaining the ignition probability is as described above. The probabilities P _{2 and 3} of the marking (1, 0) at time T 1 are P _{1, 1} × p _{1, 1 → 2, 3} + P _{1, 2} × p _{1, 1 → 2, 3} By performing the same calculation, the probability P ₃ the marking is _{present, 1,} P _{3, 2} can be calculated, P _{3, 1} and P _{3, 2} sum of that from the time T current (time T2) The probability P (B) that the operation numerical data up to is observed.

Next, how to obtain the molecule P (B | A) P (A) of the formula (19) will be described. As shown in FIG. 9 (b), assuming that the observed unobservable state exists in the initial state, when considering state transitions consistent with the observable state observed from the discrete state record 105, the marking indicated by the solid line in the initial state This is a path through which the state transitions from (0, 1) to markings (1, 0) and (1, 1) at time T1 to markings (0, 0) and (0, 1) at time T2. As in the case of FIG. 9A, ignition probabilities p _{1, 2 → 2, 3} and p _{1, 2 → 2, 4} are assigned to each solid arrow. The method for obtaining the ignition probability is as described above. The probabilities P _{2 and 3} at which the operation numerical data is observed at the marking (1, 0) at the time T1 are P _{1, 2} × p _{1, 2 → 2, and 3} . By performing the same calculation, it is possible to calculate the probabilities P _{3, 1} , P _{3, and 2 at} which the operation numerical data is observed. The probability P (B | A) P (A) that the operation numerical data from the time T to the present (time T2) is observed.

  As described above, the denominator P (B) and the numerator P (B | A) P (A) of the equation (19) are obtained, and the existence probability P (A | B) of the unobservable state at the time T is obtained. . By performing Bayesian estimation while changing the unobservable state of interest, the initial state and its existence probability are set so as not to contradict the initial state that can exist as the initial state, that is, the observable state observed from the discrete state record 105. be able to.

  Reference numeral 6 denotes a dangerous state probability calculation unit, which is a discrete state record 105 created by the discrete numerical current operation data creation unit 4, an observable state firing frequency 104 computed by the observable state firing frequency computation unit 3, and an unobservable state Based on the firing frequency 103, a reachable tree starting from the initial state set by the initial state setting unit 5 is created, and the firing probability obtained from the firing frequency 103 in the unobservable state and the firing frequency 104 in the observable state is used. Find the probability of existence of a dangerous state. In this case, the operation plan 107 may be given to calculate the temporal transition of the existence probability of the dangerous state when the operation plan 107 is executed.

  Examples of the dangerous state in the continuous casting process include those shown in Table 1, and the specific contents of each dangerous state (the attributes of the product in each stage and its state and equipment and its state) are predetermined. It has been.

  As shown in FIG. 10, based on the continuous casting model, a reachable tree is created at regular time intervals starting from the initial state set by the initial state setting unit 5. It is assumed that markings indicated by diagonal lines in the figure correspond to dangerous situations. Each solid arrow is given a firing probability. The method for obtaining the ignition probability is as described above. Therefore, the existence probability of each dangerous state can be obtained by using the existence probability and the firing probability of the initial state by the same method as that for obtaining the existence probability of the marking 8 in FIG. 5 described above.

FIG. 11 is an example of guidance information presented to the user by the operation support device for the continuous casting process, and shows a temporal transition of the probability of existence of a dangerous state when the operation is performed based on the formulated operation plan 107. The horizontal axis represents a time interval, which corresponds to a certain time interval of the reachable tree (see FIG. 10) created by the dangerous state probability calculation unit 6. The vertical axis represents the existence probability of the dangerous state calculated by the dangerous state probability calculation unit 6.
In the example of FIG. 11, the existence probability of the dangerous state “nozzle clogging” is relatively high at 2 to 4 [minutes], but the plasma heating is compared with the case where the plasma heating is not performed with the plasma heating apparatus. By doing so, it can be seen that the existence probability can be significantly lowered.
In addition, regarding the dangerous state “breakout”, it is found that the existence probability is slightly increased by plasma heating when the plasma heating device is compared with the case where plasma heating is not performed.
When there is a risk of both breakout and nozzle clogging in this way, if the risk of nozzle clogging is reduced by turning on the plasma heating device, the risk of breakout with a delay in the spatially separated mold Can be seen to rise.

  As described above, the state transition diagram can be created by estimating the existence of the unobservable state and setting the initial state so that it does not contradict the numerical data of the observable operation acquired in the current operation. The probability of falling into trouble can be evaluated in consideration of the probability of.

  In the said embodiment, although the continuous casting process was demonstrated as an example of the manufacturing process in which a continuous product flows continuously, for example, hot rolling (a steel plate flows continuously) or a petroleum plant (a liquid flows continuously) It is also applicable to.

In addition, the present invention can be applied to a manufacturing process in which a product flows. For example, a batch process performed in a chemical plant or the like is a discrete process, and a product in that process can be said to be a discrete product. In the iron and steel industry, as shown in Fig. 12, the process of transporting hot metal from a blast furnace to a steelmaking factory with a topped car, the converter process of refining hot metal into molten steel in a pot unit in the steelmaking factory-the refining process of the secondary refining process Is a discrete process, and in that process hot metal and molten steel can be said to be discrete products. If it is a refining process of a converter process in a steelmaking factory—secondary refining process, the converter process may be the first step and the second refining process may be the second step. In addition, hot metal is mixed (blended) from a plurality of topped cars to make one cup of hot metal. In that case, the combination of attributes of each topped car is transmitted to the converter.
In this case as well, using the same method as in Patent Document 1, a state transition model of product attributes and a state transition model of equipment are created for each stage in the product flow direction in the manufacturing process, and each stage Define the influence relationship in the model and perform discrete modeling. A discrete model (continuous casting model) obtained by discrete modeling of the continuous casting process is represented by, for example, a Petri net that is a directed graph. FIG. 13 shows the attributes of the product of each stage. For example, temperature, component, and quantity are defined as attributes of a product at a topped car stage. Moreover, temperature, a component, and quantity are defined as an attribute of the product in a converter stage. Moreover, temperature, a component, and quantity are defined as an attribute of the product in a secondary refining stage. Moreover, temperature, a component, and quantity are defined as an attribute of the product in a continuous casting machine stage.

A manufacturing process operation support apparatus to which the present invention is applied is realized by a computer apparatus including a CPU, a ROM, a RAM, and the like, for example.
Further, the present invention supplies software (program) for realizing the function of the operation support apparatus of the manufacturing process of the present invention to a system or apparatus via a network or various storage media, and the computer of the system or apparatus executes the program. It can also be realized by reading and executing.

  1: Database for storing actual values of observable operation numerical data 2: Discrete numerical actual value data creation unit 3: Firing frequency calculation unit for observable state 4: Discrete numerical current operation data creation unit 5: Initial state Setting unit, 6: dangerous state probability calculation unit, 101: discrete state record, 102: initial existence probability of unobservable state, 103: firing frequency of unobservable state, 104: firing frequency of unobservable state, 105: discrete state record 106: Initial existence probability of observable state 107: Operation plan

Claims (7)

A manufacturing process in which a product flows is divided into a plurality of stages in the flow direction of the product, and a state transition model of an attribute of the product and a state transition model of an equipment are created for each stage. An operation support method for supporting the operation of the manufacturing process based on a discrete model created by defining an influence relationship between the attribute state of the product and the state of the equipment,
The means for setting the initial state should be consistent with the observable operation numerical data acquired in the current operation using the state transition probability obtained from the state transition frequency of the unobservable state and the state transition frequency of the observable state. Setting an initial state by Bayesian estimation;
Based on the discrete model , the means for obtaining the existence probability creates a state transition diagram starting from the set initial state, and the state transition obtained from the state transition frequency of the unobservable state and the state transition frequency of the observable state And a step of obtaining an existence probability of a predetermined state using the probability of the manufacturing process.   The operation support method for a manufacturing process according to claim 1, wherein the manufacturing process is a manufacturing process in which a continuous product flows continuously. The state transition frequency of the unobservable state is given in advance,
The state transition frequency of the observable state having one factor in the state transition among the observable states is obtained from the actual value of the observable operation numerical data,
Among the observable states, the state transition frequency of the observable state having a plurality of factors in the state transition is the actual value of the observable operation numerical data, the initial existence probability of the observable state, and the initial existence of the unobservable state given in advance. The probability of state transition is determined from the state transition frequency of the unobservable state based on the probability and the state transition frequency of the unobservable state, and a simultaneous equation is obtained every time there is a state transition of the observable state. The operation support method of the manufacturing process according to claim 1 or 2, characterized by the above-mentioned. In the step of setting the initial state,
P (A | B) = P (B | A) P (A) / P (B)
In Bayesian estimation represented by
The posterior probability P (A | B) is a probability that an unobservable state of interest exists at the predetermined time under the condition that operation numerical data from a predetermined time to the present is observed,
P (B | A) P (A) is a probability that an unobservable state of interest exists at the predetermined time, and the operation numerical data from the predetermined time to the present is observed,
P (B) is the probability that the operation numerical data from the predetermined time to the present is observed,
When operation numerical data from the predetermined time to the present is given as evidence, pay attention to a certain unobservable state, and as the posterior probability P (A | B), the unobservable state focused on at the predetermined time The manufacturing process operation support method according to any one of claims 1 to 3, wherein the existence probability is calculated. The discrete model is represented by a directed graph in which state transition is caused by the ignition of a transition,
The transition firing probability corresponding to the state transition probability is an exponential distribution of the transition firing frequency corresponding to the state transition frequency, and the probability of the transition firing during a predetermined time is represented by a cumulative distribution function. The operation support method of a manufacturing process according to any one of claims 1 to 4, wherein A manufacturing process in which a product flows is divided into a plurality of stages in the flow direction of the product, and a state transition model of an attribute of the product and a state transition model of an equipment are created for each stage. An operation support device for supporting the operation of the manufacturing process based on a discrete model created by defining an influence relationship between the attribute state of the product and the state of the equipment,
Using the state transition frequency of the unobservable state and the state transition frequency obtained from the state transition frequency of the observable state, the initial state is set by Bayesian estimation so as not to contradict the observable operation numerical data acquired in the current operation. Means,
Based on the discrete model, create a state transition diagram starting from the set initial state, using the state transition frequency of the unobservable state and the state transition probability obtained from the state transition frequency of the observable state, An operation support apparatus for a manufacturing process, characterized by comprising means for obtaining a probability of existence of a state. A manufacturing process in which a product flows is divided into a plurality of stages in the flow direction of the product, and a state transition model of an attribute of the product and a state transition model of an equipment are created for each stage. A program for supporting the operation of the manufacturing process based on a discrete model created by defining an influence relationship between the state of the attribute of the product and the state of the equipment,
Using the state transition frequency of the unobservable state and the state transition frequency obtained from the state transition frequency of the observable state, the initial state is set by Bayesian estimation so as not to contradict the observable operation numerical data acquired in the current operation. Processing,
Based on the discrete model, a state transition diagram starting from the set initial state is created, and the state transition frequency obtained from the state transition frequency of the unobservable state and the state transition frequency of the observable state is used. A program for causing a computer to execute processing for determining the existence probability of a state.

Images