JP5463964B2 - Work content estimation apparatus and work content estimation method - Google Patents

Description

  The present invention relates to a work content estimation method and a work content estimation apparatus.

  In fields where it is necessary to carry out the determined work in accordance with the determined procedure, for example, maintenance work for elevators, confirmation of execution has been made based on written reports from workers. However, in this case, it is difficult to completely prevent work omission due to carelessness.

  In a work site where work must be performed according to a standard work manual distributed in advance, various problems may occur if the work is not performed according to the standard work manual. For example, in the maintenance work for elevators, the standard work manual is formulated with the aim of improving the efficiency of the maintenance work, so work that does not conform to the standard work manual is inefficient.

  In Patent Document 1, as means for confirming the contents of maintenance work, scheduled work operations that occur in maintenance work stored in advance, elevator operation history obtained from the elevator control device during maintenance work, and input information from a portable terminal A maintenance work evaluation means for automatically storing an operation history consisting of the above and confirming whether the maintenance work contents are incomplete by collating the scheduled work operation with the operation history and the operation history after the work is described Yes.

  In Patent Document 2, the operation information of the elevators corresponding to each maintenance work is saved, so that it is confirmed that the work has been carried out and the work time of the maintenance work is obtained. An elevator maintenance work check system is described that evaluates maintenance work performed in comparison.

JP 2005-154106 A JP 2006-151529 A

  Both prior arts describe means for associating work items with generated electronic data and confirming whether the worker has correctly performed the work. However, since the electronic data corresponding to the work item is queried from the executed work item, it is necessary to input from the outside which work item is executed. For this reason, when work other than the input work item is being performed, the fact that this work has been performed cannot be recognized. Moreover, if the input is performed manually, it becomes a burden on the user.

  The present invention has been made to solve the above-described problems, and in estimating work content from electronic data, an operator can obtain data from electronic data obtained from sensors or various electronic devices existing in the work environment. The purpose is to estimate the work performed and to automatically detect errors and omissions in the work procedure based on the result.

The present invention relates to an event extraction unit that extracts event data with time by an electronic device such as a predetermined sensor corresponding to a predetermined output condition, and a plurality of operations including time series data of a plurality of events. Using a sample of time series information, the learning unit that calculates the transition probability density function based on the semi-Markov model from the contents of each event and its stay time, and the regular work procedure that is a series of work procedures in the state and transition A work estimation unit for obtaining a posterior probability for the state from the work graph shown and time series information of actual event data is provided.

  According to the present invention, when an electronic device such as a predetermined sensor meets a predetermined output condition, an event extraction unit that extracts event data with time, and a work constituted by time-series data of a plurality of events, A learning unit that calculates a transition probability density function based on a semi-Markov model from each event stay time using statistical information calculated from a plurality of time series information samples, and a regular work procedure that is a series of work procedures By performing a work based on the regular work procedure by providing a work estimation unit that obtains a posterior probability for the state from the work graph indicated by the state and transition and the time series information of the actual event data, The work content can be estimated.

FIG. 3 is a configuration diagram of a work content estimation apparatus according to the first embodiment. An example showing extraction conditions to be extracted as event data in the work content estimation apparatus in the first embodiment. An example showing event data in the work content estimation apparatus in the first embodiment. An example of a work graph stored in the regular work procedure database 9 according to the first embodiment. An example of a work completion flag in the first embodiment in the first embodiment. 5 is a flowchart of an event extraction unit 3 in the first embodiment. 5 is a flowchart of the learning unit 6 according to the first embodiment. The figure showing matching with a semi-Markov model at the time of modeling event data in Embodiment 1. 3 is a histogram representing a distribution of transition times between two event sets in the first embodiment. 6 is an operation flowchart of estimation / diagnosis processing performed by the work estimation unit 8 and the work progress management unit 10 according to the first embodiment. 6 is a flowchart showing the operation of the candidate selection unit 101 in the first embodiment. The work graph which shows the regular work procedure in this Embodiment, The figure which also showed the present work state The figure which shows the criteria used for the candidate selection part 101 to select a work estimation candidate based on the present work state in the work graph in FIG. 6 is a flowchart of work estimation performed by the work estimation unit 8 according to the first embodiment. 6 is a flowchart of processing performed by the work determination unit 102 according to the first embodiment.

Embodiment 1 FIG.
In the first embodiment, when a worker performs a predetermined work content, electronic data is obtained from various electronic devices such as sensors that exist in the work environment, thereby the work content performed by the worker, A method for estimating the procedure will be described.

There are several methods for estimating work contents from time-series data of electronic data obtained from various electronic devices such as sensors in the work environment, which are obtained during work when a worker performs a certain work content. In this embodiment, as an example, a mode using estimation using a semi-Markov model will be described.
The semi-Markov model is a kind of Markov model and is a probability model in which the transition probability from one state to the next state is expressed as a function of the time spent in the state.
For example, when taking elevator maintenance work in mind, if one work item is taken, the order is determined to some extent, and it is possible to collate with a model that can express the order such as the state transition of the Markov model. is there.
In addition, because the maintenance work for elevators is performed by humans, there are fluctuations in the event occurrence time, but there are restrictions on the operation time such as the block sign of the event whose meaning changes depending on the length of time the button is pressed On the other hand, there are tasks that do not have time constraints, and the time variation varies from event to event.
On the other hand, the Markov model cannot express temporal blur for each event, and does not depend on time by performing state transition with the occurrence of an event as a trigger. For this reason, time constraints need to be considered separately and incidentally outside of the Markov model, whereas the temporal blur for each event that can be expressed in semi-Markov can be extended to a general transition probability function. The semi-Markov model is suitable for modeling the time-dependent blur.
Therefore, estimation using a semi-Markov model is suitable as a method for estimating work contents from time series data of electronic data from various electronic devices such as sensors existing in the work environment and its time variation.
The work content can be estimated by learning the sequence of events and the pattern of event intervals when the work is performed by the semi-Markov model, and measuring the posterior probability that can be regarded as the similarity at the time of estimation.

  FIG. 1 is a configuration diagram of a work content estimation apparatus according to the first embodiment.

  The event extraction unit 3 is connected to both an electronic device 1 that can acquire electronic data such as a sensor and a control panel, and a timer 2 that can acquire time information. Events are extracted from the electronic device 1 and the timer 2 using the acquired electronic data, and are stored in the event data storage unit 4 as event data.

FIG. 2 is an example showing extraction conditions to be extracted as event data in the work content estimation apparatus according to the first embodiment.
An event is a change in the value of electronic data such that a value of a predetermined sensor exceeds a threshold or falls within a specific range. In other words, the event data is data indicating a change in the value of the electronic data such that the value of the predetermined sensor exceeds the threshold or falls within a specific range. Also in FIG. 2, for each event ID, an electronic data series corresponding to a predetermined sensor is shown as a table for determining based on a predetermined extraction rule.
This change in the sensor value corresponds to human actions and operations such as closing the door or moving the elevator if it is applied to confirmation of elevator maintenance work. The occurrence of an event is determined according to a predetermined rule for each type.

FIG. 3 shows an example of event data in the work content estimation apparatus according to the first embodiment.
The event data stored in the event data storage unit 4 includes information on ID (type) and occurrence time.
In the data of FIG. 3, for example, an event of event ID2 that the sensor A reacts at time 12:15:00 occurs, then the switch is turned on at 12:18:32, and an event of event ID4 occurs. : The switch S is OFF at 20:02, the event ID5 event occurs, the sensor B reacts at 12:22:20, and the event ID3 event occurs.
In the above example, a case where one event ID corresponds to one occurrence time is shown, but in reality, a plurality of event IDs may occur at the same occurrence time. The work is composed of a plurality of events. Therefore, the work is estimated based on the time intervals of the plurality of event time series, and the work content cannot be determined only by detecting the event.

All work completion events and work completion events are used as special event types.
The all work completion event represents the end of work, and is generated when the maintenance work switch is turned off. The estimation / diagnosis process is terminated when all events are completed.
The work completion event is an event indicating that one work item has been completed. The work completion event is used as a signal for starting work content estimation. If there are events that always occur only at the end of a work item, they are used as work completion events. When the above condition cannot be satisfied, a work completion event can be made by inputting some sound into the microphone worn by the maintenance worker.

In the teacher data storage unit 5 in FIG. 1, the learning unit 6 stores the teacher data necessary for calculating the parameters. Necessary teacher data is a sample data collection that records in advance one work content including the same plurality of events corresponding to each event data stored in the event data storage unit 4.
For this purpose, the teacher data storage unit 5 stores a value indicated by each sensor according to work contents and a set of time zones indicated by a plurality of samples that have performed the work as one teacher data item. By giving the implementation time zones indicated by a plurality of samples and their variations, the work can be identified from the corresponding event data series by the semi-Markov model.

  The learning unit 6 first acquires a plurality of teacher data from the teacher data storage unit 5, and then acquires event data corresponding to the teacher data from the event data storage unit 4. After that, the learning unit 6 calculates and stores the estimation parameters necessary for the work estimation unit 8 to estimate the work in the estimation parameter storage unit 7. A method for calculating the estimation parameter will be described later.

  The work estimation unit 8 receives a work estimation candidate from the work progress management unit 10 and returns a work estimation result. In order to perform work estimation, information is read from the event data storage unit 4 and the estimation parameter storage unit 7. A specific work estimation method will be described later.

  The work progress management unit 10 reads the regular work procedure from the regular work procedure database 9 that stores the regular work procedure as an input, and causes the work estimation unit 8 to estimate the work procedure that has been called and executed.

The regular work procedure database 9 stores regular work procedures, which are a series of work procedures, in the form of a work graph that can be uniquely written. The work graph is defined as "a directed graph in which a node (work) is one work unit and an edge (transition) is a pre- and post-work constraint, and the same work is not executed twice".
FIG. 4 is an example of a work graph stored in the regular work procedure database 9 according to the first embodiment. The node 401 represents that a work state called work B exists, and for example, the edge 402 represents that work E can transition to work after work B.

The work progress management unit 10 compares the estimated work procedure with the regular work procedure, and if there is an error in the procedure, detects it as a work error. In addition, the work progress management unit 10 stores the work procedure and the work error information in the work procedure storage unit 121 and the work error information storage unit 122 of the work log storage unit 12.
Further, the work progress management unit 10 causes the work progress information storage unit 11 to store the work state / work candidate list in the work state storage unit 111 as temporary data during operation, and sets the work state / work completion flag to work. The completion flag storage unit 112 stores the work completion flag in the work completion flag storage unit 112, the estimation status is stored in the estimation status storage unit 113, the latest estimation result is stored in the estimation result storage unit 114, and Use.

  The work progress information storage unit 11 holds a work state, the latest estimation result, a work completion flag, and an estimation status as storage information used by the work progress management unit 10.

  The work state storage unit 111 in the work progress information storage unit 11 stores past work estimation results as the current work state. In the initial state, a special “initial work” is stored. The initial work is a work that does not have a previous work and has a work that is first performed on the work graph as a next work. However, the initial work does not correspond to the actual work and is used only by the work progress management unit 10 because it is necessary for the operation of the algorithm.

The work completion flag indicates whether the work has been performed or not performed for each work, and is stored in the work completion flag storage unit 112.
FIG. 5 shows an example of the work completion flag in the first embodiment. The work completion flag takes a value of either executed or not executed, and is not executed in the initial state.

The estimated status is obtained by counting the number of times the estimation has failed, and is stored in the estimated status storage unit 113. The estimated status is required by the candidate selection unit 101 described later. In the initial state, it is 0 times.
The estimation result storage unit 114 stores the latest estimation result estimated by the work estimation unit 8.

Next, the operation will be described with reference to FIG.
FIG. 6 is a flowchart of the event extraction unit 3 in the first embodiment. As a precondition, the event extraction rule illustrated in FIG. 2 is given in advance.
In S601, the electronic data acquired from the electronic device 1 is always observed. When any condition in the event extraction rule is satisfied in S602, the time is acquired from the timer 2 in S603, and the event is performed in S604. Are stored in the event data storage unit 4.
The event extraction unit 3 extracts events from the electronic data based on the rules. The process of the event extraction unit is continuously repeated during a period in which data can be acquired from the sensor / control panel, that is, during maintenance work.

Next, the operation of the learning unit 6 will be described. The learning unit 6 calculates a transition probability density function f of the semi-Markov model for each work as a parameter for the work estimation unit 8 to estimate the work content, and the transition probability density function f is calculated as an estimation parameter storage unit. Save to 7. Since the transition probability density function f is necessary for the estimation / diagnosis process described later, the process of the learning unit 6 needs to be executed before the estimation / diagnosis process.
FIG. 7 is a flowchart of the learning unit 6 in the first embodiment.
In S701, a plurality of pieces of information, for example, work W and the time zone required for the work W, which are teacher data, are acquired from the teacher data storage unit 5, and in S702, event data generated in the time zone is stored in the event data Obtained from part 4. Since the teacher data acquired from the teacher data storage unit 5 requires a plurality of samples even in the same work to obtain the transition probability density function f, the same work is repeatedly executed several times to obtain data in advance. It is necessary to keep.

The event data observed when the work W is performed is modeled.
FIG. 8 is a diagram showing the association with the semi-Markov model when event data is modeled in the first embodiment. Here, modeling is performed with event occurrence as state transition and event occurrence interval as stay time of each state.
At this time, it is the transition probability density function f between the working states that needs to be stored in the estimation parameter storage unit 7 as a parameter for characterizing the semi-Markov model. A semi-Markov model needs to be created for each work, and the transition probability density function f needs to be set for each combination of events.

  Various realization methods are conceivable for setting the transition probability density function of transitions, but in this implementation, the maximum entropy principle ("Fault diagnosis of event-driven systems using timed Markov models" Vol.42, no9, pp. 188-194). When the maximum entropy principle is used, the transition probability density function becomes

Referring to FIG. 8 as well, it is assumed that the average value μ of the staying time and the standard deviation σ among the samples of events from the state x to y in the operation A.
For example, ε is defined as a predetermined value when the average value μ of event staying time is more than ± 3σ. This is a predetermined value necessary not to be excluded from the estimation items in order to estimate the unusual work action / time. For example, in the above equation, the boundary of ε is ± 3σ, but the amount of variation can be adjusted to a predetermined amount depending on the contents between events.
λ (vector of λ0 to λm) is a coefficient sequence that characterizes the shape of the transition probability function f, and changes depending on the number of samples of teacher data, variation, and the like. A method for obtaining λ will be described later.
m is the order of the coefficient λ. By increasing m, it becomes possible to represent the probability distribution of the transition probability function f with a more complicated shape (such as having a plurality of peaks).
The estimation parameter storage unit 7 stores λ, average value μ, and standard deviation σ.

Returning to FIG. 7, next, in step S <b> 703, the learning unit 6 extracts a set of two events that occur successively from the acquired event data. For example, event ID 2 and event ID 4 and event ID 4 and event ID 5 in FIG. In step S704, the event occurrence intervals are totaled for each set of events, and a histogram is created.
FIG. 9 is a histogram representing the distribution of transition times, which are event intervals between two event sets in the first embodiment.
In S <b> 704, the class value wi (i = 1 to n natural number), the frequency pi (i = 1 to n natural number), the average value μ, and the standard deviation σ are calculated from the created histogram for each set of events. .

After that, the parameter λ (vector of λ0 to λm) is calculated by the Levenberg-Marquard method (see Weisstein, Eric W “Levenberg-Marqard Method From WorldWorld—A Wolf Web Resource”).
In S705, λ is initialized to an appropriate value, and then in S706, λ is updated according to the equation (2).

Here, J is an R Jacobian matrix, and I is a unit matrix.
If the value of λ converges in S707, the processing is complete.
Thereafter, λ0 is calculated by Equation 3, and λ0 to λm are stored. The parameter δ is used as the convergence condition. When the secondary Euclidean norm of the amount of change of λ is less than δ, that is, when Equation 4 is established, it is determined that the convergence has occurred.

In FIG. 7, in S708, the learning unit 6 holds Equation 4 and, after convergence, stores the λ, μ, and σ in the estimation parameter storage unit 7 in association with the type of work and the set of events.
If there is an event set for which parameters have not yet been calculated in S709, the process is performed for all event sets, and the process ends in S710.
The processing of the learning unit 6 needs to be executed for each of N works to be estimated. For example, in the example of FIG. 4, each of operations A, B, C, D, E, F, and G is executed a total of seven times.

Next, operations of work estimation / diagnosis processing performed by the work estimation unit 8 and the work progress management unit 10 will be described.
FIG. 10 is an operation flowchart of the estimation / diagnosis process performed by the work estimation unit 8 and the work progress management unit 10 according to the first embodiment. By this operation, the following object of the present invention can be achieved. That is, the work content is estimated from the event data, and whether the work content is correct is diagnosed from the estimation result.
As the pre-processing, it is assumed that the processing of the learning unit 6 has already been executed, and the parameters of the semi-Markov model are stored in the estimation parameter storage unit 7 for each work. Further, it is assumed that the event extraction unit continues to operate independently at the time of estimation / diagnosis, and the event data extracted from the electronic data is continuously stored in the event data storage unit 4.

When the operation starts in S1000 of FIG. 10, first in S1001, the candidate selection unit 101 of the work progress management unit 10 selects a work estimation candidate.
FIG. 11 is a flowchart showing the operation of the candidate selection unit 101 in the first embodiment.
The function of the candidate selection unit 101 is to select a work estimation candidate necessary for the operation of the work estimation unit 8 based on information on the work progress state and the work graph of the regular work procedure.
When the candidate selection unit 101 starts the candidate selection flow in S1100, first, in S1101, the candidate selection unit 101 acquires the current work state stored in the work state storage unit 111.
The current work state is the initial state, and the initial work is saved.
In S1102, the candidate selection unit 101 also acquires the estimated status stored in the estimated status storage unit 113.
The estimated status is set to 0 in the initial state, and increases every time work estimation fails. Moreover, it resets to 0 whenever work estimation succeeds.

In S1103, the regular work procedure is obtained from the regular work procedure database 9.
FIG. 12 is a work graph showing a regular work procedure in the present embodiment, and is a diagram showing the current work state.
In step S <b> 1104, the candidate selection unit 101 selects a work estimation candidate from the work candidate group included in the work graph of the regular work procedure in FIG. 12 and the current work state.
FIG. 13 is a diagram illustrating a determination criterion used by the candidate selection unit 101 to select a work estimation candidate based on the current work state in the work graph in FIG.
In FIG. 13, the left column has IDs of candidate groups included in the work graph of the regular work procedure, and conditions corresponding to the respective IDs are shown. A priority is given to the ID of the candidate group.
Then, the candidate selection unit 101 selects a work candidate group whose priority value matches the estimated status value from among the candidate groups shown in FIG. That is, if there is no failure in work estimation, the estimated status is 0, so in FIG. 13, work C, work D, work E, and work F in FIG. Selected.
The work corresponding to the selected work candidate group is transferred to the work estimation unit 8 as a work estimation candidate. However, at this time, if the candidate is only one task, the estimated status value is increased by 1, and the task having the next priority is added to the candidate.

The function of the candidate selection unit 101 is called every time before work estimation. Each time work estimation fails, the estimated status increases by one. Accordingly, work candidates are selected step by step according to priority each time work estimation fails.
As described above, by narrowing down the number of tasks to be referred to at one time, it is possible to improve the estimation accuracy and reduce the amount of calculation by not having to consider the estimation for extra tasks.

Next, in S1002 of FIG. 10, the work estimation unit 8 performs work estimation.
FIG. 14 is a flowchart of work estimation performed by the work estimation unit 8 according to the first embodiment. In step S <b> 1400, the work estimation unit 8 starts operation by receiving a candidate for work estimation from the candidate selection unit 101. The function of the work estimation unit 8 is to estimate which work has been performed by calculating a posteriori probability for the event data based on the estimation parameters for each work included in the received work estimation candidates. .

  When the operation of the work estimation unit 8 starts, first, in S1401, the posterior probability pM (r) is initialized to an equal value for all candidate works r (r is a natural number). The value is set so that the sum is 1. For example, in the case of FIGS. 12 and 13, if the estimated status is 0, the candidates received from the candidate selection unit 101 are four jobs C, D, E, and F, so PM (C), PM (D), PM (E) and PM (F) are each initialized to 0.25.

  Next, in S1402, one item of event data is acquired from the event data storage unit 4. The event data acquisition operation is slightly different between when the diagnosis is performed in parallel with the maintenance work during the maintenance work and when the result is diagnosed after all the maintenance work is completed as in batch processing. In both cases, one item of unacquired event data is acquired. However, if it is carried out in parallel with maintenance work, there is a possibility that unacquired event data does not exist. wait.

  In step S1403, the posterior probability pM (r) is updated for each work r in the candidate based on the acquired event and the previous event. The update formula is shown in Equation 5.

ここで、ｐＭ（ｒ）＿ｏｌｄは更新前のｐＭ（ｒ）の値であり、関数ｆは式１に示した遷移確率密度関数ｆである。また、ｅＨ＋１は、今取得したイベントの種類であり、ｅＨはその１つ前に取得したイベントの種類である。遷移確率密度関数ｆの値を算出するために必要なパラメータλ、μ、σは、推定用パラメータ保存部７から取得する。
ここでイベントデータの値から、例えば図１２、図１３の場合において、作業Ｂが終わって作業Ｃ、Ｄ、Ｅのいずれかの作業に入ったことを示した場合で、イベントデータの種類ｅＨ＋１と、前のイベントｅＨからの時間間隔から、作業Ｃ，Ｄの該当可能性が高く、作業Ｅ，Ｆの該当可能性が低い場合、遷移確率密度関数ｆの値は、作業Ｅ，Ｆよりも作業Ｃ，Ｄのほうが高くなるため、更新後の作業Ｃ，作業Ｄに対する事後確率は０．２５からそれぞれ上昇し、Ｅ，Ｆは減少するといった処理がなされる。
また、その後、次のイベントデータの種類ｅＨ＋２を抽出したときに、作業Ｃ及び作業Ｄにかかった時間によって示す遷移確率密度関数ｆが異なるため、その際に、作業Ｃ、作業Ｄに対する事後確率がさらに更新されて、最後に、作業Ｆとなるイベントデータを取得したときに、作業の推定の基礎となる事後確率が確定する。
また、イベントの組み合わせによっては、学習時に該組み合わせが存在しなかったために、パラメータが保存されていない場合がある。例えば、図８における、Ｓｔａｔｅ０からＳｔａｔｅ３などのように存在しなかった遷移の組み合わせの場合などである。
このときは遷移確率密度関数ｆの値の代わりに定数ζを用いる。ζはシステムで共通のパラメータであるが、学習時に該組み合わせが存在しない場合を想定しているので、その値は、学習時に該組み合わせが存在している場合と比べて低く設定される。

Here, pM (r) _old is the value of pM (r) before update, and the function f is the transition probability density function f shown in Equation 1. Also, eH + 1 is the type of event acquired now, and eH is the type of event acquired immediately before. Parameters λ, μ, and σ necessary for calculating the value of the transition probability density function f are acquired from the estimation parameter storage unit 7.
Here, from the value of the event data, for example, in the case of FIGS. 12 and 13, it is indicated that the work B is finished and the work C, D, or E is entered, and the event data types eH + 1 and When the possibility of work C and D is high and the possibility of work E and F is low from the time interval from the previous event eH, the value of the transition probability density function f is higher than that of work E and F. Since C and D are higher, the posterior probabilities for the updated work C and work D are increased from 0.25, and E and F are reduced.
After that, when the next event data type eH + 2 is extracted, since the transition probability density function f indicated by the time taken for the work C and the work D is different, the posterior probability for the work C and the work D is determined at that time. Further updated, and finally, when event data to be the work F is acquired, the posterior probability as the basis for the work estimation is determined.
Also, depending on the combination of events, the parameter may not be stored because the combination did not exist at the time of learning. For example, this is the case of a combination of transitions that did not exist, such as State 0 to State 3 in FIG.
At this time, a constant ζ is used instead of the value of the transition probability density function f. ζ is a parameter common to the system, but since it is assumed that the combination does not exist at the time of learning, its value is set lower than that when the combination exists at the time of learning.

After updating the posterior probability, the type of the acquired event is confirmed in S1404. In the case of a normal event that is neither a work completion event nor an all-work completion event, the process returns immediately before “acquire event data” in S1402 of FIG.
If a work completion event or all work completion events have been acquired, the work W having the maximum posterior probability is selected in S1405. When all the work completion events are acquired, the work progress management unit 10 is further notified.

When all the work completion events are acquired, the work progress management unit 10 is notified. In S1406, it is determined whether or not the posterior probability of the work W having the maximum posterior probability is larger than the threshold ξ. This threshold value ξ is a parameter, and is specified in advance as a function of the number of tasks that are candidates for task estimation.
In S1407, when the posterior probability of the work W having the maximum posterior probability is larger than the threshold ξ, the estimation is successful. In this case, the work W is stored in the estimation result storage unit 114. At the same time, the estimated status value is set to zero.
In S1408, when the posterior probability of the work W does not exceed the threshold ξ, it means that the posterior probability of any work does not exceed the threshold ξ, which is an estimation failure. The estimation failure indicates the result that it did not correspond to any work in the candidate. In this case, information indicating estimation failure is stored in the estimation result storage unit 114. At the same time, 1 is added to the estimated status value. Thereby, in the next operation of the candidate selection unit 101, the candidate group to be selected is changed, and the other candidate groups can be estimated.

In S1003 of the flowchart of the estimation / diagnosis process in FIG. 10, the estimation result (success, failure) obtained based on the flowchart of FIG. 14 is confirmed.
If the estimation result is an estimation success, the process proceeds to the process of the work determination unit 102 in S1004. If the estimation is unsuccessful, the process is repeated from the candidate selection unit 101, which is the process from S1001. However, if all the tasks (candidate group ID 4 in FIG. 13) are estimation candidates, the estimation is aborted here, and the process of the task determination unit 102, that is, S1004 is performed.

  FIG. 15 is a flowchart of processing performed by the work determination unit 102 according to the first embodiment. The work determination unit 102 determines whether the work is performed in a normal procedure. Use the work state stored in the work state storage unit 111, the estimation result stored in the estimation result storage unit 114, and the regular work procedure indicated by the work graph stored in the regular work procedure database 9. This makes it possible to check the work.

In S1501, when the operation of the work determination unit 102 is started, first, the estimation result storage unit 114 and the work state storage unit 111 are referred to, and the work and the current work state obtained as the previous estimation result are acquired. The current work state is the work that is the previous estimation result. Hereinafter, these are referred to as work W and work W ′, respectively.
In S1502, it is determined whether or not the acquired work W is an estimation failure. If the work W is an estimation failure, it can be determined in S1503 that the estimation failure has occurred even if all the above-described operations are candidates. Output. In step S1511, the estimated status 113 is set to 0, and the process ends. Then, it will progress to S1005 which is immediately after S1004 work determination of FIG.

  Next, when the acquired work W is not an estimation failure, it is checked whether the work is duplicated. In step S1504, the work completion flag stored in the work completion flag storage unit 112 is checked to check whether the item of the work W has been executed. If the item of work W of the work completion flag is already executed, the work duplication is output in S1505.

If the result of the work duplication check is not performed in S1504, a check for work omission is further performed in S1506. Specifically, it is checked whether or not there is only one edge (transition) between the work W and the work W ′ on the work graph representing the regular work procedure.
If not, in S1507, it is checked whether it is possible to reach the work W by tracing the edge from the work W ′. If it can be reached, a work omission is output in S1508, and if it cannot be reached, a procedure error is output in S1509. If there is only one edge (transition) in the work W and the work W ′, it is determined in step S1510 that the work has proceeded according to the normal procedure.

  In S1512 which is the last of the work determination process, if the work W is not an estimation failure, the work state storage unit 111 is updated to the latest estimation result. At the same time, the work W item in the work completion flag storage unit 112 is set to complete.

After the work determination, the process returns to immediately after the work determination in the flowchart of FIG. 10 of the estimation / diagnosis process, and the estimated work is output in S1005. At this time, if the estimation result stored in the estimation result storage unit 114 is an estimation failure, “unknown work” is output.
In S1006, if all the work has not been completed after the output of the estimation result, the process returns to S1001 which is the head of the flowchart to perform the next estimation.
The operation of the estimation / diagnosis process is repeatedly executed until a notification that all work completion events have occurred is received from the work estimation unit 8. If there is a notification of the occurrence of all work completion events from the work estimation unit 8, the entire process is terminated in S1007.

  Therefore, in the present embodiment, the electronic device 1 such as a predetermined sensor is configured by an event extraction unit 3 that extracts event data with time by satisfying a predetermined output condition, and time series data of a plurality of events. The learning unit 6 calculates a transition probability density function based on a semi-Markov model from the contents of individual events and their stay times using a plurality of time-series information samples, and a regular operation sequence. Work based on the regular work procedure is provided by providing the work estimation unit 8 that obtains the posterior probability for the state from the work graph indicated by the state and transition of the work procedure and the time series information of the actual event data. In this case, it is possible to estimate the work content and to detect missing work and censoring of the estimation.

  For example, when applied to elevator maintenance work confirmation, it is connected to an electronic device such as a sensor attached to the elevator or an operation control device that controls the operation of the elevator. The method and apparatus extract an electronic data change from electronic data and time information that can be acquired by an electronic device, store the type and time of occurrence as an event, time-series data of events, and A learning unit 6 that calculates and stores an estimation parameter from teacher data indicating which work was performed in a time zone; a work estimation unit 8 that estimates work content from time-series event data based on the estimation parameter; A candidate selection unit 101 that selects a candidate for a work that can be performed next from the regular procedure information by managing the work completion confirmation status and the current work state, and a work determination that determines whether the estimation result conforms to the regular procedure Part 102, and as an output, it is possible to output the procedure of the work performed by the worker and which part has an error if it does not comply with the regular procedure. Kill.

In addition, events are extracted from the electronic data during the work, and using this, the work performed by the worker is sequentially estimated to confirm whether the work is being performed according to the regular procedure. At the same time, the work procedure is output as a work history.
Electronic data is obtained from various electronic devices such as sensors that exist in the work environment, whereby the work contents and procedures performed by the worker can be estimated. By using the estimation result, it is possible to automatically detect missing maintenance work or incorrect procedures. In addition, by using the estimated procedure as a work log, it can be used as work evidence and data for improving work efficiency.

DESCRIPTION OF SYMBOLS 1 Electronic device 2 Timer 3 Event extraction part 4 Event data storage part 5 Teacher data storage part 6 Learning part 7 Estimation parameter storage part 8 Work estimation part 9 Regular work procedure database 10 Work progress management part 101 Candidate selection part 102 Work determination part 11 Work progress information storage unit 111 Work state storage unit 112 Work completion flag storage unit 113 Estimated status storage unit 114 Estimation result storage unit 12 Work log storage unit 121 Implementation work procedure storage unit 122 Work error information storage unit

Claims (18)

An event extraction unit that extracts event data with time by an output value of an electronic device capable of outputting electronic data corresponding to a predetermined output condition;
State transition in which a random variable changes from the contents of each event and its stay time using a sample of a plurality of time-series information of work composed of time-series data of a plurality of events extracted by the event extraction unit A learning unit for calculating a transition probability density function of
A transition probability density function in which the posterior probability for the state is calculated by the learning unit from the state of the regular work procedure that is a series of work procedures and the constraints before and after the state and the time series information of the event data extracted by the event extraction unit. A work content estimation apparatus comprising: a work estimation unit obtained using The learning section, claims and calculates the transition probability density function by multiplying the obtained coefficients as a solution of restricted optimization problem for service pump Le frequency distribution of each of the events residence time Item 1. The work content estimation apparatus according to Item 1.   The learning unit sets the transition probability density function to a constant value greater than 0 when the actual event stay time varies more than a predetermined value from the average of a plurality of samples with respect to the sample frequency distribution of each event stay time. The work content estimation apparatus according to claim 1, wherein the work content is estimated. It said states for said normal work procedure, the working graph Ru indicated by the transition indicating the front and rear constraints, from the work currently being estimated, then work candidates for causing demanded a posteriori probability to the working estimator 4. The work content estimation apparatus according to claim 1, further comprising a candidate selection unit that selects a work item from the work graph. 5.   Using the task candidate for causing the task estimation unit selected by the candidate selection unit to determine the posterior probability, for the task having the maximum posterior probability determined by the task estimation unit, the posterior probability is maximum. The candidate selection unit selects, from the work graph, more work candidates than the previously selected work candidates when the posterior probability of the work to be performed is equal to or less than a predetermined threshold. Described work content estimation device.   From the work graph shown in the regular work procedure and having a transition having a front-rear constraint that cannot be duplicated for the same work, the work having the maximum posterior probability obtained by the work estimation unit is the largest posterior probability. If the posterior probability of the work is less than or equal to a predetermined threshold, or exceeds the predetermined threshold, the transition between the work currently estimated from the work graph and the next work is not only one transition The work content according to any one of claims 1 to 5, further comprising a work determination unit that determines that it is a procedure error when the next work cannot be reached from the work currently estimated from the work graph. Estimating device.   From the work graph shown in the regular work procedure and having a transition having a front-rear constraint that cannot be duplicated for the same work, the work having the maximum posterior probability obtained by the work estimation unit is the largest posterior probability. The a posteriori probability of the work that exceeds the predetermined threshold is not only a single transition between the work currently estimated from the work graph and the next work, but is currently estimated from the work graph. 6. The work content estimation apparatus according to claim 1, further comprising a work determination unit that determines that the work is skipped when the work can reach the next work.   From the work graph shown in the regular work procedure and having a transition having a front-rear constraint that cannot be duplicated for the same work, the work having the maximum posterior probability obtained by the work estimation unit is the largest posterior probability. The work posterior probability exceeds the predetermined threshold, but the work determination unit is provided to determine that the work is currently duplicated when the work currently estimated from the work graph and the next work overlap. The work content estimation apparatus according to claim 1, wherein the work content estimation apparatus is a work content estimation apparatus.   From the work graph shown in the regular work procedure and having a transition having a front-rear constraint that cannot be duplicated for the same work, the work having the maximum posterior probability obtained by the work estimation unit is the largest posterior probability. If the posterior probability of a given work exceeds a predetermined threshold and there is only one transition between the work currently estimated from the work graph and the next work, it is determined that the work is a normal work. The work content estimation apparatus according to claim 1, further comprising a work determination unit. An event extraction step in which the event extraction unit extracts the event data with time because the output value of the electronic device that can output the electronic data meets a predetermined output condition;
Using a sample of multiple time series information of work consisting of time series data of multiple events, the transition probability density function of the state transition in which the random variable changes from the content of each event and its stay time , A learning step calculated by the learning unit ;
And the condition and its front and rear constraints for normal operation procedure is a series of working procedures, and a time-series information of the actual event data, by using the transition probability density function of the posterior probability is calculated by the learning unit with respect to the state, working A work content estimation method comprising: a work estimation step obtained by an estimation unit . 11. The transition probability density function according to claim 10, wherein, in the learning step, the transition probability density function is obtained by multiplying a coefficient obtained as a solution of the limited optimization problem by a sample frequency distribution of each event stay time. Described work content estimation method.   In the learning step, when the actual event stay time varies from the average of a plurality of samples to the sample frequency distribution of each event stay time, the transition probability density function is set to a constant value greater than zero. The work content estimation method according to claim 10 or 11, characterized in that: It said states for said normal work procedure, the working graph Ru indicated by the transition indicating the front and rear constraints, from the work currently being estimated, then work candidates for causing demanded a posteriori probability to the working estimator 13. The work content estimation method according to claim 10, further comprising a candidate selection step of selecting from the work graph.   Using the task candidate for causing the task estimation unit selected by the candidate selection step to determine the posterior probability, the task having the maximum posterior probability determined by the task estimation unit is the maximum posterior probability. The candidate selection unit selects, from the work graph, more work candidates than the previously selected work candidates when the posterior probability of the work to be performed is equal to or less than a predetermined threshold. Described work content estimation method.   From the work graph shown in the regular work procedure and having a transition having a front-rear constraint that cannot be duplicated for the same work, the work having the maximum posterior probability obtained by the work estimation step has the maximum posterior probability. If the posterior probability of the work is less than or equal to a predetermined threshold, or exceeds the predetermined threshold, the transition between the work currently estimated from the work graph and the next work is not only one transition The work content estimation according to claim 10, further comprising: a work determination step that determines that a procedure error occurs when the next work cannot be reached from the work currently estimated from the work graph. Method.   From the work graph shown in the regular work procedure and having a transition having a front-rear constraint that cannot be duplicated for the same work, for work that has the maximum posterior probability obtained by the work estimation step that maximizes the posterior probability, The posterior probability of the work exceeds a predetermined threshold, but the current work is estimated from the work graph as well as one transition between the work currently estimated from the work graph and the next work. 15. The work content estimation method according to claim 10, further comprising a work determination step for determining that the work is skipped when the work can reach the next work.   The posterior probability is maximized with respect to the work having the maximum posterior probability obtained by the work estimation step from the work graph having transitions having front-rear constraints that cannot be duplicated for the same work as indicated in the regular work procedure. The posterior probability of work exceeds a predetermined threshold, but when the work currently estimated from the work graph overlaps with the next work, the work posterior probability includes a work determination step for determining that the work is duplicated. The work content estimation method according to claim 10, wherein the work content is estimated.   From the work graph shown in the regular work procedure and having a transition having a front-rear constraint that cannot be duplicated for the same work, the work having the maximum posterior probability obtained by the work estimation step has the maximum posterior probability. If the posterior probability of a given work exceeds a predetermined threshold and there is only one transition between the work currently estimated from the work graph and the next work, it is determined that the work is a normal work. The work content estimation method according to claim 10, further comprising a work determination step.

Images