JP7072531B2 - Anomaly detection device and anomaly detection method - Google Patents

Description

The present invention relates to an abnormality detection device and an abnormality detection method.

In Patent Document 1, "in a DTW notation in which the standard time series data calculated in the standard time series data calculation step is arranged on one axis and the target time series data acquired in the abnormality detection target data acquisition step is arranged on the other axis. An abnormality is defined as a distance calculation step for calculating the distance between the standard time series data and the target time series data, and a distance calculated in the distance calculation step larger than a predetermined reference in the time series data. An abnormality detection method including an abnormality data point detection step to be detected ”is disclosed.

Japanese Unexamined Patent Publication No. 2009-245228

In the above-mentioned Patent Document 1, in order to detect an abnormality in time-series data, the system expands and contracts the standard time-series data and the target time-series data by DTW (Dynamic Time Warping). The degree of abnormality is measured by finding the matching points between the two, calculating the distance, and summing up, and determining whether or not it is abnormal. This makes it possible to make comparisons without regard to time lag. On the other hand, it is difficult to detect an abnormality in the required time (delay of a certain operation, etc.).

An object of the present invention is to provide a technique capable of detecting anomalies on time-series data in consideration of required time anomalies such as delay and advancement.

The present application includes a plurality of means for solving at least a part of the above problems, and examples thereof are as follows.

One aspect of the present invention is an abnormality detection device, which comprises a standard time-series model numerical parameter in which a reference time-series data is modeled, and a standard transition time model numerical parameter in which the duration of each state is modeled. State estimation by the standard time series model numerical parameters for the model storage unit to be stored, the time series data input unit that accepts the input of the time series data to be the target of abnormality detection, and the time series data to be the target of the abnormality detection. The state estimation unit that estimates the state variable of each data point, the standard transition time model numerical parameter, and the state variable of each data point estimated by the state estimation unit are used to determine each data point. Of these, an abnormality degree calculation unit that calculates the degree of abnormality of the time required for transition from a predetermined state to another state, and an abnormality degree output unit that outputs the above abnormality degree as a graph as a delay degree for each elapsed time. Be prepared.

INDUSTRIAL APPLICABILITY According to the present invention, it is possible to provide a technique capable of detecting an abnormality in consideration of a required time abnormality on time-series data.

Issues, configurations and effects other than those described above will be clarified by the following description of the embodiments.

It is a figure which shows the outline of the example of the required time abnormality detection part in 1st Embodiment. It is a figure which shows the example of the initial probability table of the standard time series model numerical parameter storage part. It is a figure which shows the example of the output probability table of the standard time series model numerical parameter storage part. It is a figure which shows the example of the transition probability table of the standard time series model numerical parameter storage part. It is a figure which shows the example of the transition time table of the standard transition time model numerical parameter storage part. It is a figure which shows the example of the abnormality degree memory table of the abnormality degree calculation part. It is a figure which shows the example of the effect of a smoothing process. It is a figure which shows the example of the abnormality degree output screen (peak list). It is a figure which shows the example of the abnormality degree output screen (delay time point estimation). It is a figure which shows the example of the abnormality degree output screen (delay degree degree). It is a figure which shows the outline of the example of the required time abnormality detection part in the 2nd Embodiment. It is a figure which shows the outline of the example of the required time abnormality detection part in 3rd Embodiment. It is a figure which shows the example of the result of the matching operation. It is a figure which shows the example of the required time abnormality detection system in 4th Embodiment. It is a figure which shows the configuration example of the work abnormality detection system in 4th Embodiment. It is a figure which shows the example of the required time abnormality detection system in the 5th Embodiment. It is a figure which shows the example of the skeleton recognition processing in the 5th Embodiment. It is a figure which shows the example of the required time abnormality detection system in the 6th Embodiment. It is a figure which shows the structural example of the abnormality degree output part in 7th Embodiment. It is a figure which shows the example of the abnormality degree output screen (work control).

In the following embodiments, when necessary for convenience, the description will be divided into a plurality of sections or embodiments, but unless otherwise specified, they are not unrelated to each other, and one is the other. It is related to some or all modifications, details, supplementary explanations, etc.

Further, in the following embodiments, when the number of elements (including the number, numerical value, quantity, range, etc.) is referred to, except when explicitly stated or when the number is clearly limited to a specific number in principle. , The number is not limited to the specific number, and may be more than or less than the specific number.

Furthermore, it goes without saying that, in the following embodiments, the components (including element steps and the like) are not necessarily essential unless otherwise specified or clearly considered to be essential in principle. stomach.

Similarly, in the following embodiments, when the shape, positional relationship, etc. of the constituent elements are referred to, the shapes, etc. It shall include those that are close to or similar to. This also applies to the above numerical values and ranges.

Further, in all the drawings for explaining the embodiment, the same members are in principle the same reference numerals, and the repeated description thereof will be omitted. Hereinafter, each embodiment of the present invention will be described with reference to the drawings.

DTW (Dynamic Time Warping) can be expressed by a superordinate concept of matching operation. In the present embodiment, the method of calculating the matching point by expanding and contracting the time series data with time is defined as the matching operation. The matching operation involves time expansion and contraction, and even if the rate of change in the standard time series data and the target time series data is different, the matching point can be obtained.

In general, standard time series data and matching operations can be represented by the superordinate concepts of standard time series model and state estimation, respectively. Here, the state estimation is a calculation method in which the state variables corresponding to each data point are estimated by using the time series data as an input and the numerical parameters, and the numerical parameters and the calculation method used are used as a standard time series model. The matching operation can be considered as a kind of state estimation in the sense that each data point of the standard time series data is treated as a different state and each data point of the given target time series data is matched with one of them. However, when an abnormality such as a delay is observed in the transition time, this cannot be detected as an abnormality.

Therefore, in the embodiment of the present invention, a standard transition time model expressing the standard transition time of each state is newly introduced. For the section from any data point A to data point B for calculating the degree of abnormality of the target time series data, refer to the state included in the target section obtained from the result of state estimation for the target time series data and the standard transition time model. From the transition time distribution of each state obtained in the above, the transition time distribution for the section or its representative value is calculated. The degree of nonconformity of the interval transition time from the data point A to the data point B in the target time series data to the interval transition time distribution, or the difference between the transition time and the representative value of the interval transition time distribution is calculated and used as the degree of abnormality.

[First Embodiment] FIG. 1 is a diagram showing an outline of an example of a required time abnormality detection unit in the first embodiment. Here, it is assumed that a plurality of time-series data for which input is received by the time-series data input unit 103 are connected in order of elapsed time from the time when the data points are obtained. The variable representing the time when the data point is obtained is defined as the time stamp, and the variable representing the value measured at the data point is defined as the data value. Data values and time stamps are arranged in chronological order as a data value series and a time stamp series, and it is possible to retain and express time series data by these two. However, it is not essential to use a time stamp series in the representation of time series data, and if the data points are sampled according to a certain rule, it is possible to keep the sampling rule instead of the time stamp series. It can be calculated from. The data values include not only scalars but also vectors and images.

Further, it is assumed that the model storage unit 102 is preset with information obtained by modeling a state of a predetermined single or a plurality of reference time-series data in advance. The preset method can be realized by the method shown in the second embodiment.

FIG. 2 is a diagram showing an example of an initial probability table of the standard time series model numerical parameter storage unit. The initial probability table 1021 is a table showing the probabilities of the initial state of the standard time series model in a single time series data. The initial probability table 1021 includes the state 1021A and the initial probability 1021B in association with each other. That is, it can be said that the probability that the state is the initial state is stored for each state.

FIG. 3 is a diagram showing an example of an output probability table of the standard time series model numerical parameter storage unit. The output probability table 1022 is a table showing the output probabilities of the standard time series model in a single time series data. The output probability table 1022 includes the state 1022A, the mean value 1022B, and the variance 1022C in association with each other. That is, it can be said that the distribution of values that can be taken in each state is stored for each state.

FIG. 4 is a diagram showing an example of a transition probability table of the standard time series model numerical parameter storage unit. The transition probability table 1023 is a table showing the probability of transition between the states of the standard time series model in a single time series data. The transition probability table 1023 includes the probability of transition for the combination of the original state 1023A and the transition destination state 1023B in association with each other. That is, it can be said that the probability that the state transitions to another state is stored for each state.

FIG. 5 is a diagram showing an example of a transition time table of the standard transition time model numerical parameter storage unit. The transition time table 1024 is a table that models and shows the distribution of the duration of a standard time series model for each state in a single time series data. The transition time table 1024 includes the state 1024A, the mean value 1024B, and the variance 1024C in association with each other. That is, it can be said that the distribution of possible values for the duration of the state is stored for each state.

The required time abnormality detection unit 108 receives the time-series data from the time-series data input unit 103, and passes the abnormality degree information to the abnormality degree output unit 104. Time-series data to be detected as an abnormality is input to the time-series data input unit 103. The required time abnormality detection unit 108 performs a matching operation with the model information preset in the model storage unit 102 for the time series data, identifies the abnormality degree, and passes it to the abnormality degree output unit 104. The abnormality degree output unit 104 performs output processing such as displaying using the abnormality degree information or recording the abnormality degree information in a storage area. At this time, the abnormality degree output unit 104 may output a warning when the abnormality degree is larger than a specific reference.

The required time abnormality detection unit 108 includes an abnormality degree calculation unit 100, a state estimation unit 101, a model storage unit 102, and a time stamp acquisition unit 105.

The time stamp acquisition unit 105 acquires the time stamp series of the target time series data received from the time series data input unit 103 and sends it to the abnormality degree calculation unit 100.

The state estimation unit 101 contains standard time-series model numerical parameters read from the model storage unit 102 (including the initial probability table 1021 in FIG. 2, the output probability table 1022 in FIG. 3, and the transition probability table 1023 in FIG. 4). For the time-series data received by the time-series data input unit 103, state estimation is performed by a standard time-series model, state variables of each data point are estimated, and the estimated state variables are sent to the abnormality degree calculation unit 100. send. The state estimation unit 101 uses a state estimation method corresponding to the standard time series model to estimate the state variables. For example, when a hidden Markov model is used as a standard time series model, the state estimation unit 101 performs state estimation using a Viterbi algorithm.

Further, the state estimation unit 101 integrates the results of state estimation using a plurality of standard time series model numerical parameters to estimate state variables. For example, the state estimation unit 101 may estimate the mean value or the median value of a plurality of estimated state variables as state variables, or may obtain the mean value and the variance and estimate them as probabilities.

The anomaly degree calculation unit 100 reads out the standard transition time model numerical parameters from the model storage unit 102, the state variables of each data point sent from the state estimation unit 101, and the time stamp series received from the time stamp acquisition unit 105. , Is calculated and stored in the abnormality degree storage table 1001.

The abnormality degree calculation unit 100 time stamps the transition time T1 required for transitioning from the data point A to the data point B for the target section from the data point A to the data point B for calculating the abnormality degree of the target time series data. Obtain from the series and acquire the state included in the target section from the result of state estimation for the target time series data. Further, the anomaly degree calculation unit 100 is a section which is a transition time distribution over the entire section from the transition time distribution of each state obtained by referring to the standard transition time model or its representative value (mean value, mode value, etc.). Derivation or approximation of the transition time distribution or its representative value T2. The abnormality degree calculation unit 100 calculates the difference between the transition time T1 and the representative value T2, or the transition time T1 and the interval transition time distribution as the abnormality degree.

As a method for approximating the interval transition time distribution, the anomaly degree calculation unit 100 adopts a sampling method. As a method for approximating the representative value, the abnormality degree calculation unit 100 adopts a method of taking a cumulative sum when the representative value is an average value.

As a method of calculating the difference between the interval transition time distribution and the transition time T1, the degree to which the transition time T1 fits the interval transition time distribution is evaluated by the likelihood, etc., and the reciprocal or positive / negative inversion is abnormal. The method used as a degree is adopted.

Further, as a method of calculating the difference between the transition time T1 and the representative value T2, for example, a method of dividing the transition time T1 by the representative value T2 is adopted. In this method, when the degree of abnormality, which is the quotient of the result of division, becomes larger than "1", is there a delay in the transition from data point A to data point B in the target time series data, or is there a delay in the transition? It indicates that an event different from the data is likely to be inserted between the data point A and the data point B. In other words, it is highly possible that the work operation was delayed or another work operation was performed during the work operation. On the contrary, when the degree of abnormality becomes smaller than "1", the change from the data point A to the data point B is accelerated in the target time series data, or the event existing in the standard time series data is data from the data point A. It indicates that there is a high possibility that the data is missing between the points B. In other words, it is highly possible that the work operation was performed earlier than usual, or that the work operation was omitted.

Further, in the above, an example of calculating the degree of abnormality (delay degree or advancement degree) between arbitrary two points (between data point A and data point B) is shown, but the present invention is not limited to this, and the abnormality degree calculation unit 100 may perform the abnormality degree calculation unit 100. It is also possible to calculate the degree of abnormality as a series by repeating the calculation of the degree of abnormality for two points at sufficiently short time intervals on the target time series data.

FIG. 6 is a diagram showing an example of an abnormality degree storage table of the abnormality degree calculation unit. The anomaly degree storage table 1001 includes a time point 1001A, which is information for specifying a time point, and an abnormality degree 1001B associated with the time point 1001A.

Further, the abnormality degree calculation unit 100 may perform post-processing such as smoothing on the calculated abnormality degree series. As the smoothing algorithm, a known method such as a moving average method is adopted.

FIG. 7 is a diagram showing an example of the effect of the smoothing process. In FIG. 7, the line before smoothing is represented by a dotted line, and the line after smoothing is represented by a solid line. For example, when the time-series data is affected by noise or the like, the variation in the degree of abnormality becomes large. Even in such a case, by performing smoothing, it is possible to obtain information on the degree of abnormality with suppressed variation as shown in FIG. 7.

The abnormality degree calculation unit 100 stores the calculated abnormality degree series or the smoothed information in the abnormality degree storage table 1001 and passes it to the abnormality degree output unit 104.

FIG. 8 is a diagram showing an example of an abnormality degree output screen (peak list). The abnormality degree output screen (peak list) 1041 is screen information created by the abnormality degree output unit 104 by graphing a series of abnormality degrees or information obtained by smoothing the series. The abnormality degree output screen (peak list) 1041 has a delay degree display area 1041A for displaying a delay degree graph showing the magnitude of the delay degree (abnormality degree) generated for each product, and a delay degree display area 1041A for each product in descending order of the degree of delay. It includes a peak list 1041B that displays the time when the peak value of the degree is taken, and a confirmation button 1041C that accepts an instruction to end the screen display. This screen allows the user to compare the peak delays of each product. Not limited to this, the abnormality degree output unit 104 may create and output the following output screen.

FIG. 9 is a diagram showing an example of an abnormality degree output screen (delay time point estimation). The abnormality degree output screen (delay time point estimation) 1042 is screen information created by the abnormality degree output unit 104 by graphing a series of abnormality degrees or information obtained by smoothing the series. On the abnormality degree output screen (delay time estimation) 1042, a delay degree display area 1042A for displaying a delay degree graph showing the magnitude of the delay degree (abnormality degree) generated in the product, and a time point when a particularly large delay degree is detected. (Peak) includes a comment 1042B that displays the occurrence of a delay, and a confirmation button 1042C that accepts an instruction to end the screen display. From this screen, the user can know the time when a large delay occurs in the target product and get a hint for improving the work content.

FIG. 10 is a diagram showing an example of an abnormality degree output screen (delay degree). The abnormality degree output screen (delay degree) 1043 is screen information created by the abnormality degree output unit 104 by graphing a series of abnormality degrees or information obtained by smoothing the series. On the abnormality degree output screen (delay degree) 1043, a delay degree display area 1043A for displaying a delay degree graph showing the magnitude of the delay degree (abnormality degree) generated in the product, a time zone in which the delay degree is detected, and the like thereof are displayed. A comment 1043B for displaying the amount of delay in time and a confirmation button 1043C for receiving an instruction to end the screen display are included. From this screen, the user can know the time when the delay occurs continuously for the target product and can obtain a hint for improving the work content.

In order to create the abnormality degree output screen (delay degree) 1043, it is necessary to specify the delay amount in the time zone in which the delay degree is detected. In this specifying process, the abnormality degree calculation unit 100 specifies the delay amount by calculating the difference (T1-T2) between the transition time T1 and the representative value T2. By doing so, the delay amount can be specified, and the abnormality degree output screen (delay degree degree) 1043 can be created. In addition, the user can also grasp the amount of delay by creating a graph for the amount of abnormality and using geometric information (specifically, an area indicating the amount of delay) on the graph.

[Second Embodiment] FIG. 11 is a diagram showing an outline of an example of a required time abnormality detection unit in the second embodiment. The required time abnormality detection unit 108 ′ according to the second embodiment is basically the same as the required time abnormality detection unit 108 according to the first embodiment, but there are some differences. Hereinafter, the difference from the required time abnormality detection unit according to the first embodiment will be mainly described.

In the second embodiment, the standard model stored in the model storage unit 102 can be learned from a single or a plurality of standard time series data obtained from the time series data input unit 103.

The required time abnormality detection unit 108 ′ according to the second embodiment further includes a standard time series model estimation unit 106 and a standard transition time model estimation unit 107.

The standard time series data input from the time series data input unit 103 is passed to the standard time series model estimation unit 106, the state estimation unit 101, and the time stamp acquisition unit 105. The time stamp acquisition unit 105 calculates or acquires a time stamp series and passes it to the standard transition time model estimation unit 107.

The standard time-series model estimation unit 106 estimates a single or a plurality of numerical parameters of the standard time-series model using the passed single or a plurality of standard time-series data and stores them in the model storage unit 102. The learning of the standard time series model involves estimating numerical parameters using AI (artificial intelligence), etc., such as a machine learning framework such as a hidden Markov model or a probabilistic automaton. For example, when a hidden Markov model is used, the initial probability, the transition probability, and the output probability are estimated as numerical parameters, and the model storage unit 102 uses the estimated numerical parameters (initial probability table 1021, output probability table 1022, transition probability table 1023). ) Is stored for each standard time series data.

The state estimation unit 101 uses the numerical parameters of the standard time series data read from the model storage unit 102 to estimate the state using the standard time series model for the set of standard time series data input from the time series data input unit 103. To estimate the state variable of each data point of the time series data. The state estimation unit 101 passes the estimated state variables to the standard transition time model estimation unit 107. For the estimation of state variables, the state estimation method corresponding to the standard time series model is used. For example, in the case of a hidden Markov model, the state can be estimated using the Viterbi algorithm. Further, when a plurality of numerical parameters of standard time series data are used, since a plurality of states are estimated, the state estimation unit 101 may obtain the distribution thereof.

The standard transition time model estimation unit 107 receives a state variable of each data point from the state estimation unit 101, and receives a time stamp of each point from the time stamp acquisition unit 105. By associating the state with the time stamp, the standard transition time model estimation unit 107 calculates the transition time from the start of a specific state to the transition to the next state in the target time series data.

Then, the standard transition time model estimation unit 107 calculates the transition time for all the pairs in the plurality of standard transition time models of the state variable and the time stamp of each data point, so that the transition time for each state is calculated. Is obtained and stored as a standard transition time model (transition time table 1024) of the model storage unit 102. The transition time distribution is approximated by the standard transition time model estimation unit 107 with a Gaussian distribution or the like, and the parameters (mean, variance, etc.) are stored as representative values. Not limited to this, all calculated transition times may be saved as they are.

By doing so, it is possible to learn a single or a plurality of standard models stored in the model storage unit 102 by using the standard time series data obtained from the time series data input unit 103.

[Third Embodiment] FIG. 12 is a diagram showing an outline of an example of a required time abnormality detection unit in the third embodiment. The required time abnormality detection unit 108 ″ according to the third embodiment is basically the same as the required time abnormality detection unit 108 according to the second embodiment, but there are some differences. Hereinafter, the difference from the required time abnormality detection unit according to the second embodiment will be mainly described.

In the third embodiment, the required time abnormality detection unit 108 ″ receives a single (series) standard time series data and uses it as a standard model, so that the standard time series model estimation unit 106 and the standard transition time model estimation unit 106 ′. This is an example in which 107 is unnecessary.

The required time abnormality detection unit 108 ″ according to the third embodiment includes an abnormality degree calculation unit 100, a state estimation unit 101, a model storage unit 102, and a time stamp acquisition unit 105.

The time series data input unit 103 accepts input of a single standard time series data. The time-series data input unit 103 sends the data value series in which the data values are arranged in time series to the model storage unit 102, and the time stamp series to the time stamp acquisition unit 105.

The time stamp acquisition unit 105 acquires the time stamp series of the input standard time series data and passes it to the model storage unit 102.

Then, the model storage unit 102 stores the data value series of the standard time series data passed from the time series data input unit 103 and the time stamp series passed from the time stamp acquisition unit 105. This is the learning flow for a single standard model. Next, the flow at the time of calculating the degree of abnormality will be described.

When the time-series data input unit 103 receives the target time-series data to be detected as an abnormality, the time-series data input unit 103 sends the data value series to the state estimation unit 101, and the time stamp series or the information necessary for its calculation. It is sent to the time stamp acquisition unit 105 respectively. The time stamp acquisition unit 105 acquires or calculates the time stamp series of the target time series data and passes it to the abnormality degree calculation unit 100.

The state estimation unit 101 performs a matching operation on the data value series of the passed target time series data and the data value series of the standard time series data read from the model storage unit 102, and the result of the matching calculation is obtained by the abnormality degree calculation unit 100. Hand over to.

FIG. 13 is a diagram showing an example of the result of the matching operation. This result is not limited to the third embodiment, and is the same in all embodiments from the first embodiment to the seventh embodiment described later. The matching points of the data value series of the target time series data shown by the solid line and the data value series of the standard time series data are connected by diagonal lines. It is the matching operation that specifies the points to be connected, and the state estimation unit 101 adopts and specifies a known algorithm such as DTW.

The calculation of the degree of abnormality by the abnormality degree calculation unit 100 is basically the same as that of the first embodiment and the second embodiment. In the configuration of the third embodiment, since a single standard time series data is used as the numerical parameters of the standard time series model, the standard time series model estimation is performed by modeling using the plurality of standard time series data shown in the second embodiment. The unit 106 becomes unnecessary. In addition, since each data point of the standard time series data is treated as a different state and the state estimation is realized by associating each data point of the target time series by the matching operation, the standard transition time of each state is a single value. Yes, it is the same as the difference between adjacent time stamps in the time stamp series. Therefore, the time stamp series may be stored as it is as the standard transition time model, and the standard transition time model estimation unit 107 is also unnecessary. Therefore, the required time abnormality detection unit 108 ″ according to the third embodiment can be simplified in its configuration.

[Fourth Embodiment] FIG. 14 is a diagram showing an example of a required time abnormality detection system in the fourth embodiment. In the fourth embodiment, an example in which a required time abnormality detection unit is applied to a work abnormality detection system 205 that detects an abnormality in manual work by a wearable sensor (a detachable sensor that acquires the amount of change in position). be.

It is assumed that the worker 201 attaches the wearable sensor 202 to the body and repeatedly performs the work. The sensor data of the movement of the worker 201 measured by the wearable sensor 202 is sent to the information processing terminal 203 via wireless communication. As the wearable sensor 202 for measuring the movement, a sensor such as an acceleration sensor or a gyro sensor that acquires the amount of change in position can be used. When the information processing terminal 203 receives the sensor data at the sensor receiving unit 204, the information processing terminal 203 inputs the sensor data to the work abnormality detection system 205.

FIG. 15 is a diagram showing a configuration example of the work abnormality detection system according to the fourth embodiment. Sensor data (time-series data) is passed from the sensor receiving unit 204 to the feature quantity extraction unit 206, and the feature quantity extraction unit 206 converts the sensor data into time-series data composed of feature quantities suitable for abnormality detection. For example, if the sensor data is missing, the feature amount extraction unit 206 performs complementary processing and removes noise contained in the sensor data with a filter such as a low-pass filter, so that the accuracy of abnormality detection can be expected to be improved. .. Further, the feature amount extraction unit 206 obtains the feature amount by Fourier transforming the sensor data.

The time-series data from which the feature amount has been extracted by the feature amount extraction unit 206 is passed to the series section cutout unit 207, and based on the control signal from the cutout control unit 210, the section corresponding to one operation in the repetitive work. Is cut out and handed over to the required time abnormality detection unit 108.

The required time abnormality detection unit 108 calculates the degree of abnormality with respect to the received time-series data as shown in the first, second, and third embodiments described above, and outputs the calculation result to the abnormality degree output unit 104. send. The abnormality degree output unit 104 records the abnormality degree and displays it on a display or the like. By doing so, for example, by checking the degree of abnormality displayed on the screen, it becomes possible to determine whether the work is being performed normally. If the degree of abnormality is greater than a specific standard, a warning may be output to assist workers and supervisors in making it easier to deal with.

[Fifth Embodiment] FIG. 16 is a diagram showing an example of a required time abnormality detection system in the fifth embodiment. In the fifth embodiment, the required time abnormality detection unit is applied to the work abnormality detection system 205 that detects abnormalities in manual work by a camera.

It is assumed that the worker 301 is performing the work repeatedly. The video data of the worker 301 taken by the camera 302 is sent to the information processing terminal 303 via wired communication or wireless communication. The camera 302 is not limited to a camera that captures visible light, and may capture invisible light such as infrared rays, and any camera that arranges still images in chronological order to form a moving image is sufficient.

When the information processing terminal 303 receives the moving image data from the camera 302, the information processing terminal 303 inputs the sensor data to the work abnormality detection system 205.

The information processing terminal 303 can regard the moving image data as time-series data in which frame data are arranged in chronological order. The work abnormality detection system 205 according to the present embodiment is basically the same as the work abnormality detection system 205 according to the fourth embodiment, but there are some differences. Hereinafter, the differences from the work abnormality detection system 205 according to the fourth embodiment will be mainly described.

The feature amount extraction unit 206 extracts the joint position of the subject in the moving image data as the feature amount. Specifically, the feature amount extraction unit 206 analyzes the moving image data and performs skeleton recognition. For example, using a skeleton recognition algorithm, the feature amount extraction unit 206 recognizes a skeleton.

FIG. 17 is a diagram showing an example of the skeleton recognition process in the fifth embodiment. In the skeleton recognition process, skeleton recognition is performed for each frame 310 of the moving image data of the worker 301, and the joint position of the worker is extracted (frame 311). By using the joint position as a feature amount, it is possible to convert it into time-series data that reflects the movement of the worker 301, and it becomes easy to detect an abnormality.

The processing after the time-series data extracted by the feature amount is passed to the series section cutting unit 207 is the same as that of the fourth embodiment.

According to the fifth embodiment, for example, by checking the degree of abnormality displayed on the screen, it is possible to determine whether the work is being performed at a normal speed. If the degree of abnormality is larger than a specific standard, a warning may be output to assist the operator, the supervisor, or the like to check the operation with the image taken by the camera 302 and facilitate the countermeasure.

[Sixth Embodiment] FIG. 18 is a diagram showing an example of a required time abnormality detection system in the sixth embodiment. In the sixth embodiment, the required time abnormality detection unit is applied to the work abnormality detection system 205 that detects the abnormality of the machine tool work by the current sensor.

The machine tool 501 is a machine that performs cutting, and the magnitude of the current value of the spindle power line corresponds to the operating state of the machine tool 501 such as cutting, idling, and stopping. The current sensor 502 sends the measured current value to the information processing terminal 503 via wired communication or wireless communication.

When the information processing terminal 503 receives the time-series data of the current value from the current sensor 502, the information processing terminal 503 inputs the sensor data to the work abnormality detection system 205.

The work abnormality detection system 205 according to the present embodiment is basically the same as the work abnormality detection system 205 according to the fourth embodiment, but there are some differences. Hereinafter, the differences from the work abnormality detection system 205 according to the fourth embodiment will be mainly described.

The feature amount extraction unit 206 converts the sensor data into time-series data composed of feature amounts suitable for abnormality detection for the time-series data of the current value. For example, by removing the noise contained in the sensor data with a filter such as a low-pass filter, the accuracy of abnormality detection can be expected to be improved. For example, if the sensor data is missing, the feature amount extraction unit 206 performs complementary processing and removes noise contained in the sensor data with a filter such as a low-pass filter, so that the accuracy of abnormality detection can be expected to be improved. .. Further, the feature amount extraction unit 206 obtains the feature amount by Fourier transforming the sensor data.

The processing after the time-series data extracted by the feature amount is passed to the series section cutting unit 207 is the same as that of the fourth embodiment.

According to the sixth embodiment, for example, by checking the degree of abnormality displayed on the screen, it is possible to determine whether or not the machine tool 501 is performing the work operation at a normal speed. If the degree of abnormality is greater than a specific standard, a warning may be output to assist in facilitating measures such as repairing machine tools and arranging state maintenance work.

[Seventh Embodiment] FIG. 19 is a diagram showing a configuration example of an abnormality degree output unit in the seventh embodiment. In the seventh embodiment, the output of the degree of anomaly is expanded. The seventh embodiment has basically the same configuration as the required time abnormality detection system according to any one of the first to sixth embodiments, but there are some differences. The difference will be described below.

As shown in FIG. 19, in the seventh embodiment, the anomaly degree output unit 104 holds an identifier of an event that occurs on the standard time series data, and time information about the time and duration of the event. The standard identification information 402 is provided, and the standard identification information 402 is connected to the abnormality degree display unit 401. The standard identification information 402 includes, for example, in the example of work in a factory as shown in the fourth to sixth embodiments, the order of work in a series of work and the standard work time of each are predetermined. In many cases, information is stored in advance based on this.

When the information about the abnormality degree of the specific time is passed from the required time abnormality detection unit 108, the abnormality degree display unit 401 refers to the standard identification information 402 and searches for the identifier of the event occurring at the corresponding time on the standard time series data. Then, create a display screen by associating the degree of abnormality with the identifier.

FIG. 20 is a diagram showing an example of an abnormality degree output screen (work control). The abnormality degree output screen (work control) 1044 is screen information created by the abnormality degree output unit 104 by graphing a series of abnormality degrees or information obtained by smoothing the series. On the abnormality degree output screen (work control) 1044, a delay degree display area 1044A for displaying a delay degree graph showing the magnitude of the delay degree (abnormality degree) generated in the product and an identifier display for displaying a work identifier for each time point are displayed. An area 1044B and a confirmation button 1044C for receiving an instruction to end the screen display are included. From this screen, the user can know the work at the time when a large delay occurs for the target product and get a hint for improving the work content.

Here, the time information for the identifier held by the standard identification information 402 is only for standard time, and if a time abnormality occurs frequently in the target time series data, the error may be so large that it cannot be ignored. Therefore, the abnormality degree display unit 401 may calculate the time lag from the standard case in the target time series data based on the abnormality degree, correct the time information, and then use the identifier. Specifically, in the abnormality degree calculation unit 100, the section from the start time to the specific elapsed time is divided into local sections, and the abnormality degree of each local section AB is obtained by the difference between T1 and T2 (definition of T1 and T2). Is the same as in the first embodiment), by calculating this cumulative value, it is possible to calculate the time lag with respect to the standard time up to a specific time, and use this to correct the time information before using it for display. Can be done. The above is the seventh embodiment.

The present invention is not limited to the above-described embodiment, and includes various modifications. For example, the above-described embodiment has been described in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to the one including all the described configurations. Further, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment.

Further, it is possible to add / delete / replace a part of the configuration of each embodiment with another configuration. Further, each of the above configurations, functions, processing units, processing means and the like may be realized by hardware by designing a part or all of them by, for example, an integrated circuit. Further, each of the above configurations, functions, and the like may be realized by software by the processor interpreting and executing a program that realizes each function. Information such as programs, tables, and files that realize each function is stored in a memory, a hard disk, a recording device such as an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, or a DVD (Digital Versaille Disc). Can be placed.

In addition, the control lines and information lines indicate what is considered necessary for explanation, and do not necessarily indicate all the control lines and information lines in the product. In practice, it can be considered that almost all configurations are interconnected.

The present invention is not limited to the abnormality detection unit, the abnormality detection device, and the abnormality detection method, and can be provided in various forms such as an abnormality detection system, a computer-readable program, and an image processing circuit.

100 ... Abnormality calculation unit, 101 ... State estimation unit, 102 ... Model storage unit, 103 ... Time series data input unit, 104 ... Abnormality output unit, 105 ... Time stamp acquisition unit , 108 ... Required time abnormality detection unit.

Claims (13)

A model storage unit that stores standard time-series model numerical parameters that model the reference time-series data and standard transition time model numerical parameters that model the duration of each state.
A time-series data input unit that accepts input of time-series data that is the target of abnormality detection,
A state estimation unit that estimates the state of the time-series data to be detected by the standard time-series model numerical parameters and estimates the state variables of each data point.
Abnormality of time required for transition from a predetermined state to another state of each data point using the standard transition time model numerical parameter and the state variable of each data point estimated by the state estimation unit. Anomalous degree calculation unit that calculates the degree, and
An abnormality degree output unit that graphs and outputs the abnormality degree as a delay degree for each elapsed time, and
Anomaly detection device characterized by being equipped with. The abnormality detection device according to claim 1.
It is equipped with a time stamp acquisition unit that acquires a time stamp from the time series data that is the target of the abnormality detection.
The anomaly degree calculation unit makes a transition from each state to another state by using the standard transition time model numerical parameter, the state variable of each data point estimated by the state estimation unit, and the time stamp. Calculate the degree of abnormality of the required time,
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
A plurality of the standard time series model numerical parameters and the standard transition time model numerical parameters are stored in the model storage unit.
The state estimation unit estimates the state variables by integrating the results of state estimation using a plurality of standard time series model numerical parameters.
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
A standard time series model estimation unit that estimates single or multiple numerical parameters of a standard time series model using the passed single or multiple standard time series data and stores them in the model storage unit as the standard time series model numerical parameters.
Anomaly detection device characterized by being equipped with. The abnormality detection device according to claim 1.
A standard that calculates the transition time from the start of a specific state to the transition to the next state in the passed single or multiple standard time series data and stores it in the model storage unit as the standard transition time model numerical parameter. Transition time model estimation unit,
Anomaly detection device characterized by being equipped with. The abnormality detection device according to claim 1.
The time-series data input unit receives the input of time-series data output from the sensor that acquires the amount of change in position.
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
Equipped with a feature amount extraction unit that analyzes video data and recognizes the skeleton.
The time-series data input unit receives the input of video data output from the camera and receives the input.
The state estimation unit estimates the state using the change in the joint position of the skeleton obtained as a result of analyzing the moving image data by the feature amount extraction unit as time series data.
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
The time-series data input unit receives an input of time-series data output from a sensor that acquires a current value.
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
The abnormality degree output unit stores in advance standard identification information including an identifier of an event that occurs on the reference time-series data and time information about the time and duration of the event.
In the graphing, the identifier of the event corresponding to the predetermined state is associated.
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
The abnormality degree output unit indicates an elapsed time in which the peak value of the delay degree becomes large in association with the graph.
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
The abnormality degree output unit indicates a comment of the occurrence of the delay at the position of the elapsed time when the peak value of the delay degree of the graph becomes large.
Anomaly detection device characterized by this. The abnormality detection device according to claim 1.
The anomaly degree output unit creates the graph for the amount of the anomaly degree, and indicates the amount of the delay degree by the geometric information on the graph.
Anomaly detection device characterized by this. It is an abnormality detection method using an information processing device.
The information processing device is
It is equipped with a model storage unit and a control unit that store standard time-series model numerical parameters that model the reference time-series data and standard transition time model numerical parameters that model the duration of each state.
The control unit
A time-series data input step that accepts input of time-series data that is the target of anomaly detection,
A state estimation step for estimating the state variable of each data point by performing state estimation with the standard time series model numerical parameters for the time series data to be the target of the abnormality detection, and
Abnormality of time required for transition from a predetermined state to another state of each data point using the standard transition time model numerical parameter and the state variable of each data point estimated in the state estimation step. Abnormality calculation step to calculate the degree and
An abnormality degree output step that graphs and outputs the abnormality degree as a delay degree for each elapsed time, and
Anomaly detection method characterized by carrying out.

Images