JP2020052676A - State monitor and wind power generator using the same - Google Patents

Abstract

To provide a state monitor capable of discriminating spike-like noise on an early stage, and a wind power generator including the same.SOLUTION: A state monitor 1 includes a measurement device 20, a measurement data acquisition unit 30 that acquires data from the measurement device 20, a measurement data memory unit 40 that preserves data obtained from the measurement data acquisition unit 30, and a measurement data arithmetic unit 50 that computes data obtained from the measurement data memory unit 40. When the data obtained from the measurement data memory unit 40 does not exceed a threshold designated on the basis of the data preserved in the measurement data memory unit 40, the measurement data arithmetic unit 50 does not predict or calculate time-sequential data. When the data obtained from the measurement data memory unit 40 exceeds the threshold, the measurement data arithmetic unit predicts or calculates the time-sequential data using a time-sequential prediction model.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a condition monitoring device and a wind power generation device using the same.

  2. Description of the Related Art A state monitoring device that monitors the state of a power generator such as a wind power generator, an industrial machine, a railway, a vehicle, and the like is known. Such a state monitoring device monitors, for example, vibration, temperature, and the like of a monitoring target, and notifies the occurrence of abnormality if the vibration is abnormal.

  In such a state monitoring device, spike noise may occur in the observation signal. If spike noise is present in the observation signal, it may be erroneously detected that the monitoring target is abnormal.

  Japanese Patent Laying-Open No. 2008-140100 (Patent Literature 1) discloses an information processing apparatus that determines spike noise in time-series data.

JP 2008-140100 A

Takuya Yoshimura and 2 others, "Proposal of a solar power generation prediction model based on a memory-based particle filter", IPSJ Research Report, Research Report Intelligent System (ICS), 2013-ICS-172, No. 8, p. 1 -7, November 12, 2013

  In the time series data obtained by the state monitoring system, the acquired value may temporarily increase due to the influence of noise or the like, and it is necessary to determine whether the change in the time series data is noise.

  In the method disclosed in Japanese Patent Application Laid-Open No. 2008-140100, when a time-series information processing device obtains an actual measurement value corresponding to the predicted value predicted using past data, Then, it is determined whether or not the noise is caused by comparing the predicted value with the predicted value. If it is noise, the measured value is replaced with normal data.

  However, the above method requires an actual measurement value at a time earlier than the acquisition time of the spike noise candidate data, and cannot determine whether the spike noise is spike noise at the time of acquiring the candidate data. For example, time-series data acquired by a state monitoring device or the like installed in a wind power generator is not always operating at all times, and data may not be acquired periodically. For this reason, it is not known when the next data after the acquired data will be obtained. Therefore, in this method, since the determination process is not performed until an actual measurement value at a time earlier than the acquisition time of the spike noise candidate data is obtained, there is a problem that the response is delayed. Even when data can be acquired periodically, there is a similar problem when the data acquisition interval is long.

  An object of the present invention is to solve such a problem, and an object of the present invention is to provide a state monitoring device capable of determining spike-like noise at an early stage and a wind power generation device including the same.

  The present disclosure relates to a state monitoring device. The state monitoring device includes a measurement device, a measurement data collection unit that collects data from the measurement device, a measurement data storage unit that stores data obtained from the measurement data collection unit, and a data obtained from the measurement data storage unit. A measurement data calculation unit for performing calculation. The measurement data calculation unit performs a prediction calculation of the time-series data when the data obtained from the measurement data storage unit does not exceed a threshold set based on the data stored in the measurement data storage unit. Instead, when the data obtained from the measurement data storage unit exceeds the threshold value, the prediction calculation of the time-series data is performed using the time-series prediction model.

  Preferably, the state monitoring device further includes a diagnosis unit that diagnoses a state of the monitoring target based on the data collected by the measurement data collection unit. The diagnostic unit switches the processing based on the result of the prediction calculation executed by the measurement data calculation unit.

  More preferably, the measurement data calculation unit is limited to the case where the data obtained from the measurement data storage unit exceeds the threshold, sets the data that exceeds the threshold as an initial value, and The prediction calculation is executed with the time at which the data was measured as the initial time.

  Preferably, the time series prediction model is a memory-based particle filter. The measurement data calculation unit refers to the data stored in the measurement data storage unit, calculates a probability density distribution after a predetermined time from the initial time using a memory-based particle filter, and initializes the probability density distribution based on the calculated probability density distribution. A predicted value after a predetermined time from the time is calculated.

Here, the predicted value is a calculation result of point estimation or section estimation.
More preferably, the measurement data calculation unit obtains a similarity between the initial value and each data stored in the measurement data storage unit. The measurement data calculation unit searches for a time having a high degree of similarity, and predicts time-series data after a predetermined time from the initial time by referring to data obtained after a predetermined time from the time obtained by the search.

  In another aspect, the present disclosure relates to a wind power generator including the state monitoring device according to any one of the above.

  The state monitoring device according to the present disclosure, when detecting acquired data that exceeds a threshold, even if there is no actual value in the future after that time, determines whether or not the data is likely to temporarily increase. Can be notified together with a predicted value after a predetermined time from the initial time. For example, when spike noise candidate data is detected, the device can be diagnosed based on the prediction result after a predetermined time even if there is no future actual value after that time.

It is a schematic diagram of a configuration of a state monitoring device of the present embodiment. It is a figure for explaining an outline of a spike noise judgment processing of a study example. FIG. 4 is a diagram for describing an outline of a spike noise determination process according to the present embodiment. 2 is a flowchart for explaining processing of a measurement data calculation unit in FIG. 1. 3 is a flowchart for explaining detailed processing of a measurement data calculation unit and processing of a diagnosis unit in FIG. 1. It is a figure showing an example of prediction of time series data predicted by a measurement data calculation part. FIG. 7 is a diagram illustrating a typical example of a probability density distribution result calculated in a prediction calculation for determining a prediction value D (T0 + Δt) in FIG. 6. It is a figure showing an example of a probability density distribution result (when accuracy of a prediction value is low when standard deviation is large). It is a figure which shows an example of a probability density distribution result (when distribution has multimodality and it is not appropriate to set to one prediction value). It is a figure showing the example of categorization of a prediction result. It is a conceptual diagram when the conditions which start time series prediction calculation are satisfy | filled. FIG. 4 is a conceptual diagram of initial particle generation during a memory-based particle filter prediction calculation. It is a conceptual diagram of the particle likelihood calculation process at the time of memory-based particle filter prediction calculation. FIG. 11 is a diagram illustrating an output example of a probability density distribution result at T0 + Δt after a predetermined time from an initial time. It is a figure which shows the example of output of the prediction value (point estimation value) in T0 + (DELTA) t after predetermined time from an initial time.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings below, the same or corresponding portions have the same reference characters allotted, and description thereof will not be repeated.

  FIG. 1 is a schematic configuration diagram of a state monitoring device according to the present embodiment. The monitoring target of the state monitoring device of the present embodiment is the wind power generator 10 in the example of FIG. The monitoring target may be other than the wind power generator 10. The state monitoring device according to the present embodiment may be, for example, a device that monitors a power generation device other than a wind power generation device, an industrial machine, a railway, a vehicle, and the like.

  The state monitoring device 1 of the present embodiment includes a measurement device 20, a measurement data collection unit 30, a measurement data storage unit 40, a measurement data calculation unit 50, and a diagnosis unit 60. The measuring device 20 is a vibration sensor, a temperature sensor, another sensor, or the like attached to the equipment of the wind turbine generator 10.

  The measurement data collection unit 30 collects data from the measurement device 20. The measurement data collection unit 30 is configured to acquire data from the measurement device 20 regularly or irregularly. Note that the acquisition may be either automatic acquisition or manual acquisition.

  The measurement data storage unit 40 stores data obtained from the measurement data collection unit 30. The measurement data storage unit 40 stores the data acquired from the measurement data collection unit 30 in a database such as a memory or a hard disk.

  The measurement data calculation unit 50 calculates data obtained from the measurement data storage unit 40. The measurement data calculation unit 50 reads the data stored in the measurement data storage unit 40, predicts time-series data as needed, and determines spike noise. If the data obtained from the measurement data storage unit 40 does not exceed the threshold set based on the data stored in the measurement data storage unit 40, the measurement data calculation unit 50 predicts the time-series data. When the data obtained from the measurement data storage unit 40 exceeds the threshold without performing the calculation, the prediction calculation of the time-series data is performed using the time-series prediction model.

  The diagnosis unit 60 switches the content of the processing based on the result of the prediction calculation executed by the measurement data calculation unit 50. The diagnosing unit 60 performs an abnormality diagnosis of the monitoring target corresponding to the determination result of the measurement data calculating unit 50.

  The first feature of the state monitoring device 1 shown in the present embodiment is that only when peculiar data is detected, such as when spike noise candidate data is detected, the data acquisition time is set to the initial time, and the acquired value is set to the initial time. As an initial value, a prediction calculation after a predetermined time from the initial time is performed.

  When it is desired to predict whether or not unusual data will continue, performing a predictive calculation at other normal times (every time data is acquired or periodically) is a measure of processor usage and memory capacity involved in the calculation. Waste of resources. Further, when dealing with transient unique data such as spike noise data, it is not necessary to perform long-term prediction. By predicting only the state at the scheduled measurement time next to the data in which the peculiar data is detected, it is possible to determine whether the data is noise or not, depending on whether the state continues or returns to the previous state. In the present embodiment, the start condition and the prediction period of the time-series prediction calculation regarding the prediction target are considered in this way.

  The second characteristic of the state monitoring device 1 according to the present embodiment is that, based on the result of predicting the time-series data at a predetermined time after the initial time, the spike-like noise is generated without an actual measurement value after the current time. Is quantitatively shown. The state monitoring device 1 according to the present embodiment, in a technology for determining whether or not a spike-like noise is temporarily increased, determines only from data up to an initial time and a future predicted value calculated from a time-series prediction calculation. I do. Thereby, the subsequent diagnosis can be performed without waiting for obtaining the measured value after the initial time.

  FIG. 2 is a diagram for explaining the outline of the spike noise determination process of the study example. In the spike noise determination process of the study example (the method of Japanese Patent Application Laid-Open No. 2008-140100), a predicted value after time N + 1 is obtained using data up to time N. When the measured value after time N + 1 is obtained, it is determined whether the measured value is within the prediction error range. If the measured value after time N + 1 is within the prediction error range, it is determined to be noise, and if it is outside the prediction error range, it is determined that it is not noise.

  That is, in this method, when a predicted value obtained by predicting the data at time N + 1 using the past data up to time N is determined to be noise when an actual measurement value after the predicted time (time N + 1) is obtained. Is determined.

  In the method of the study example, the actual measurement value at time N + 1 or later is necessary after the prediction calculation in order to determine whether the actual measurement value at time N + 1 is spike noise. Therefore, at time N at which an actually measured value cannot be obtained after time N + 1, it cannot be determined whether the noise is spike noise.

  FIG. 3 is a diagram for explaining an outline of the spike noise determination processing according to the present embodiment. In order to clarify the difference from the study example, a description will be given with reference to FIG. 3 which is created so as to be easily compared with FIG.

  In the method of the present embodiment, at time N, a predicted value at time N + 1 is obtained by the memory-based particle filter method using data up to time N. That is, the time series prediction model to be applied is different from the method of the study example. The memory-based particle filter method models the probability of the past state reappearing from the history of the state of the observation target that appeared in the past, and uses this to generate a prior distribution of the future state by random sampling from the history. Method.

  Next, it is determined whether or not the actually measured value at the time N is a spike noise from only the predicted value at the time N + 1. This point is significantly different from the study example.

  In the determination method according to the present embodiment, as shown in the upper part of FIG. 3, when the predicted value at time N + 1 is approximately the same as that before time N-1, there is a high possibility that the noise is spike noise. Is determined. On the other hand, as shown in the lower part of FIG. 3, when the predicted value at time N + 1 is much larger than the value up to time N-1, it is determined that there is a high possibility that the noise is not spike noise.

  Note that this determination method is performed under the precondition that spike noise is transient and is rarely measured continuously.

  In comparison with FIG. 2, in the determination method of the present embodiment shown in FIG. 3, prediction of the spike noise may be performed by performing prediction calculation, and subsequent measured values after time N + 1 are unnecessary. In other words, the determination method of the present embodiment can determine whether or not the actual measurement value at time N is spike-like noise at the time N at which the actual measurement value cannot be obtained after time N + 1 if there is a predicted value at time N + 1.

  The noise determination method described in the present embodiment uses a time-series prediction model such as a memory-based particle filter method at the time of acquiring data exceeding a threshold in detecting spike noise or the like in the time-series data. Thus, the prediction is performed a predetermined time after the initial time, the prediction result is output, and the noise determination can be performed early based on the prediction result.

  FIG. 4 is a flowchart for explaining the processing of the measurement data calculation unit in FIG. The processing of the flowchart shown in FIG. 4 is executed by the measurement data calculation unit 50 of FIG. The measurement data calculation unit 50 sequentially executes a time-series data generation process (Step S1), a prediction target data detection process (Step S2), a prediction calculation process (Step S4), and an output process (Step S5).

  In the time series data generation processing executed in step S1, the measurement data calculation unit 50 reads the data accumulated in the measurement data storage unit 40 and generates time series data.

  Subsequently, in the prediction target data detection process executed in step S2, a threshold value is set, and the data acquired in step S1 is compared with the threshold value. Then, in step S3, the measurement data calculation unit 50 detects whether the acquired data has exceeded a set threshold value. Data exceeding the threshold value becomes prediction target data for performing noise determination.

  As a method of setting the threshold value, a method of setting statistically based on data in the measurement data storage unit 40, a method of setting empirically, and the like are considered. For example, when the candidate data detection threshold is set to Yth indicated by a broken line in FIG. 6 described later, the latest actual value D (T0) is a data candidate for spike noise.

  In the prediction calculation process executed in step S4, a prediction result after a predetermined time from the initial time is calculated using a time-series prediction model, using the data candidates detected in step S3 as initial values. The “predetermined time” is determined by the user from the acquisition interval of the measurement data and the like.

  As an example of a time-series prediction model to be used, a memory-based particle filter (see Non-Patent Document 1) can be used. Note that a predicted value can be obtained by a time-series prediction model or a statistical method other than the memory-based particle filter.

  In particular, applying the memory-based particle filter method as a time-series prediction calculation model of the state monitoring system has the following advantages.

  For example, in the vibration data acquired by the wind power generator condition monitoring system, unique data unique to each wind turbine may be frequently measured depending on the external environment or the internal environment of the wind turbine. Therefore, a model that predicts the near future based on past data measured by the windmill is preferable. In the case where the past data to be referred to is small, such as immediately after the start of the operation of the windmill, and the number of similar cases in the past is small, the predicted value includes uncertainty. Therefore, a model that outputs the uncertainty is preferable. These problems can be solved by applying the memory-based particle filter method to the condition monitoring system.

  Further, the similarity calculation process in the memory-based particle filter method described in the present embodiment is based on past similar data for calculating a future predicted value in the state monitoring system in consideration of wind turbine-specific past data. Is a process for searching for. With this process, the peculiarity of the data itself can be reflected in the predicted value as in the case of each wind turbine measurement data.

  FIG. 5 is a flowchart for explaining the detailed processing of the measurement data calculation unit and the processing of the diagnosis unit in FIG. Steps S2 to S5 in FIG. 5 are processes executed by the measurement data calculation unit 50, and correspond to steps S2 to S5 in FIG. 4, respectively. Step S6 in FIG. 5 is a process executed by the diagnosis unit 60.

  First, in step S1, the measurement data calculation unit 50 reads data accumulated in the measurement data storage unit 40 and generates time-series data. Subsequently, a prediction data detection process in step S2 is executed. First, in step S21, the measurement data calculation unit 50 calculates a prediction calculation execution threshold Yth based on the latest time-series data, and passes the prediction calculation execution threshold Yth to the next process.

  In step S22, the measurement data calculation unit 50 detects a prediction target data candidate. Specifically, the measurement data calculation unit 50 determines whether the value of the latest time-series data exceeds the prediction calculation execution threshold Yth. Then, the determination result (TRUE or FALSE) of whether or not the threshold value is exceeded is transferred to the next processing.

  Subsequently, in step S3, when the determination result of exceeding the threshold value is FALSE, the processing of this flowchart ends, and neither the noise determination nor the diagnosis is performed. On the other hand, when the determination result of exceeding the threshold value is TRUE in Step S3, the measurement data calculation unit 50 determines that the prediction target data candidate is the prediction target data D (T0), and proceeds to Step S41 and the subsequent steps. .

  In step S41, the measurement data calculation unit 50 performs an initial setting. In the initial setting, the latest time-series data and the initial conditions (such as the number of particles) of the memory-based particle filter are used as input data. The measurement data calculation unit 50 arranges many particles around the candidate data and generates an initial distribution. As a result, initial particle data is obtained (corresponding to P1 in FIG. 12 described later).

  Subsequently, in step S42, a likelihood (Yudo) calculation process is performed. In the likelihood calculation, a search for similar cases and determination of reference data are performed. The input data in the likelihood calculation includes the latest time-series data and the number of reference data (up to the highest similarity). The contents of the likelihood calculation processing include: 1) calculating a similarity with past time-series data; 2) setting a time having a higher similarity as a reference time Tref; and 3) a predetermined time after the reference time Tref. Weighting the initial particles with the value at the next measurement timing Tref + Δt as a reference value D (Tref + Δt). As a result of the likelihood calculation, the likelihood of each particle is obtained.

  Here, the similarity is calculated based on the Euclidean distance between the initial value and each time-series data. It is also possible to adjust the type of data to be referenced and the range of data to be referenced.

  Subsequently, in step S43, a weight calculation process is executed. The weight is the weight of a particle. The input data of the weight calculation process is the likelihood of each particle. In the weight calculation process, the likelihood of each particle is normalized, and the weight of the particle is calculated. The measurement data calculation unit 50 plots the weights and generates a probability density distribution. As a result of the weight calculation process, a weight value of each particle and a probability density distribution are obtained.

  That is, in the weight calculation process, the measurement data calculation unit 50 calculates the weight of each particle of the initial distribution using data after a predetermined time from the reference time.

  Thereafter, in step S44, a predicted value calculation process is executed. The input data of the prediction value calculation processing is the weight value of each particle. In the prediction value calculation process, the measurement data calculation unit 50 calculates a point estimation value from the weight value of each particle. As one of the calculation methods, there is a method of using a point estimation value as a weighted average of weight values of respective particles. As a result of the prediction value calculation process, a point estimation value a predetermined time after the initial time is obtained.

  That is, in the predicted value calculation process, the measured data calculation unit 50 calculates the probability density distribution at a predetermined time after the initial time from the weight of each particle, and based on the calculated probability density distribution at a predetermined time after the initial time (point estimation). Value).

  Subsequently, in step S51, a predicted value category setting process is executed. The input data of the prediction value classification setting process is time-series data up to the latest. The measurement data calculation unit 50 determines a delimiter value for categorizing predicted values from the latest time-series data. As a result of the predicted value section setting process, a delimiter value for the predicted value section is obtained as output data.

  Subsequently, in step S52, a predicted value processing is performed. The input data of the predicted value processing is a point estimated value after a predetermined time from the initial time, a delimiter value for predicted value division, and a probability density distribution. The measurement data calculation unit 50 compares the point estimation value after a predetermined time from the initial time with the delimiter value for predictive value classification, and categorizes the predictive value (level division). However, if the reliability of the point estimation value is low from the probability density distribution, no categorization is performed. As a result of the predicted value processing, a predicted value level a predetermined time after the initial time is obtained as output data.

  In the prediction calculation of the memory-based particle filter according to the present embodiment, the initial time is set to the time T0 at which data satisfying the condition of the prediction data candidate is measured, and the initial value is set to the value Y (T0) of the data.

  The following is the detailed contents of the diagnosis processing S6 executed by the diagnosis unit 60 in FIG. Referring to FIG. 5, first, in step S61, diagnostic section 60 determines whether or not the initial value is predicted to be temporarily increased.

  When the initial value is predicted to be a temporary increase (YES in S61), in step S62, the diagnostic unit 60 executes a diagnostic process for spike noise determination. The diagnostic processing for spike noise determination uses the latest time series data and its related data as input data. Then, the diagnosis unit 60 performs a noise diagnosis on the latest time-series data. If the condition monitoring device is attached to the wind power generator, the latest data of the vibration data is analyzed, and it is diagnosed whether the vibration data contains noise.

  On the other hand, when the initial value is not predicted to be a temporary increase (NO in S61), in step S63, the diagnostic unit 60 executes a diagnostic process for non-spike noise determination. The diagnosis processing for non-spike noise determination uses the latest time series data and its related data as input data. In the case of the condition monitoring device attached to the wind power generator, the diagnosis unit 60 performs a vibration analysis or the like on the latest data.

  If it is determined in step S3 that the latest time-series data is not a prediction target candidate and if a diagnosis has been made in step S62 or S63, then the processing of this flowchart ends.

  Hereinafter, the processing performed by the state monitoring device according to the present embodiment will be described with reference to specific observation examples.

FIG. 6 is a diagram illustrating a prediction example of the time-series data predicted by the measurement data calculation unit. The time-series data in FIG. 6 is calculated from the vibration data of the wind turbine generator. The vertical axis indicates the effective value [m / s ² ] of the vibration, and the horizontal axis indicates the observation date and time. Although FIG. 6 illustrates the time-series vibration data measured by the wind power generator, the diagnosis method described in the present embodiment can be applied to any time-series data in which spike noise may be mixed.

  The prediction result D (T0 + Δt) using the weighted average value of the particles as the predicted value is indicated by a mark in FIG. It is predicted that the prediction result D (T0 + Δt) drops sharply to the actual value before D (T0), which has risen sharply, and that the time-series data changes spike-like.

  FIG. 7 is a diagram illustrating a typical example of a probability density distribution result calculated in the prediction calculation for determining the predicted value D (T0 + Δt) in FIG. The vertical axis in FIG. 7 indicates the probability density. The horizontal axis in FIG. 7 indicates the value of the time-series data similar to the vertical axis in FIG.

  The predicted value D (T0 + Δt) in FIG. 6 is a weighted average of the probability values plotted in FIG.

  Each of the plots in FIG. 7 is data for each particle, and the horizontal axis is determined by the value at which the initial particle was generated. FIG. 7 shows the result of the prediction calculation, in which the value at which the initial particles are generated (= the range that the predicted value can take) is set to 0.0 to 0.07. That is, the range that the predicted value can take is the range plotted on the horizontal axis in FIG. As shown in FIG. 12, which will be described later, the range in which the predicted value can be obtained differs depending on the initial value and the method of generating the initial particles.

  In the output processing in step S6, the result of the prediction calculation is processed and output. If the time series model calculates a probability density distribution as a calculation result like a particle filter, the prediction value is combined with the probability density distribution. In this case, the range that the predicted value can take can be quantitatively indicated.

  Although the graph (probability density distribution) shown in FIG. 7 is a typical example, one predicted value can be output regardless of the shape of the graph. As an output result that supports whether or not the predicted value is reliable, the probability density distribution shown in FIGS. 8 and 9 can be used.

  FIG. 8 is a diagram illustrating an example of a probability density distribution result (when the standard deviation is large and the accuracy of the predicted value is low). FIG. 9 is a diagram illustrating an example of a probability density distribution result (in the case where the distribution has multimodality and it is not appropriate to determine one predicted value).

  The measurement data calculation unit 50 only presents the graph when the observer can determine the reliability of the predicted value based on the shape of the graph of the probability density distribution (can determine whether it is appropriate to determine one predicted value). Perform On the other hand, when the observer cannot judge by the graph alone, the measurement data calculation unit 50 outputs a statistical value separately so that the observer can judge.

  In the case where it is not appropriate to determine one predicted value, the measurement data calculation unit 50 sets the determination result of the spike noise to “unknown” because the reliability is lacking.

  In particular, when the standard deviation is large and the accuracy of the predicted value is low (FIG. 8), or when the distribution is multimodal (FIG. 9) and it is not appropriate to determine one predicted value, the output of the probability density distribution is used. Is a useful notice for the observer.

  FIG. 10 is a diagram illustrating an example of categorizing prediction results. As shown in FIG. 10, by defining in advance the range in which the actual value can be taken, the predicted value can be processed and the result of categorization can be output. In FIG. 10, the data is divided into two sections, Level1 and Level2, but the number of sections and the method of dividing the sections are not limited to this.

  By outputting such a prediction result, diagnosis based on the prediction value can be performed by the diagnosis unit 60 without waiting for a future actual value after the data D (T0) detected in the prediction target data detection processing in step S2. Can be implemented.

  The diagnosis unit 60 performs a diagnosis according to the predicted value output from the measurement data calculation unit 50 as shown in FIG.

  For example, in the case of a condition monitoring diagnosis attached to a wind turbine generator, as shown in FIG. 10, when the predicted values are categorized and the predicted values are within a range indicated by Level1, the initial values are temporarily set. It is estimated that the possibility of the rise is high, and a spike noise determination diagnosis such as a noise determination process is performed.

  On the other hand, if the predicted value is in the range of Level 2, it is estimated that the increased value is likely to continue from the initial value, and diagnosis for non-spike noise determination such as vibration analysis of the equipment is performed.

  Of course, the diagnostic method or the diagnostic processing to be performed may be switched based on the prediction value or the probability density distribution result that is not categorized.

  FIG. 11 is a conceptual diagram when the condition for starting the time-series prediction calculation is satisfied. The start condition of the prediction calculation will be described with reference to FIG. In the prediction calculation of the memory-based particle filter according to the present embodiment, the initial time is set to the time T0 at which data satisfying the condition is measured, and the initial value is set to the value Y (T0) of the data.

  The threshold value Yth is obtained based on the data stored in the measurement data storage unit 40. It is assumed that the initial value Y (T0) of data D (T0) exceeds threshold value Yth.

  In this case, since the initial value Y (T0) exceeds the threshold value Yth, the condition for performing the prediction calculation is satisfied.

  FIG. 12 is a conceptual diagram of initial particle generation at the time of memory-based particle filter prediction calculation. FIG. 12 shows an example of a diagram in which initial particles are dispersed at the time of memory-based particle filter prediction calculation. Particles of the same size are scattered around data D (T0) giving the initial value Y (T0) to generate initial particles.

  FIG. 13 is a conceptual diagram of a particle likelihood calculation process at the time of memory-based particle filter prediction calculation.

  In the particle likelihood calculation processing, similarity is calculated first. The similarity is calculated by calculating the similarity between the initial value Y (T0) of the data D (T0) and each piece of past time-series data by the method described in Equation 14 of Non-Patent Document 1, and the similarity is high. The time is set as a reference time Tref. The reference time Tref in FIG. 13 is the time when the data D (Tref) most similar to the initial value is measured.

  With reference to the data Y (Tref + Δt) at a time Tref + Δt that is a predetermined time after the reference time Tref, the initial particles are weighted, and a weight distribution P2 of T0 + Δt after a predetermined time from the initial time is set.

  FIG. 14 is a diagram illustrating an output example of a probability density distribution result at T0 + Δt after a predetermined time from the initial time. The vertical axis of FIG. 14 represents the value obtained by normalizing the likelihood of the particle shown in FIG. 13, and the horizontal axis represents the predicted value (for example, the effective value) of the particle.

  FIG. 15 is a diagram illustrating an output example of a predicted value (point estimation value) at T0 + Δt after a predetermined time from the initial time. In FIG. 15, this predicted value is indicated by D (T0 + Δt). A predicted value (point estimation value) D (T0 + Δt) at T0 + Δt after a predetermined time from the initial time is calculated based on the probability density distribution after a predetermined time from the initial time shown in FIG. Note that this point estimation value is a weighted average value of the particles.

  A method using a memory-based particle filter, which is a time-series prediction model used in the present embodiment, will be summarized with reference to FIGS. The measurement data calculation unit 50 refers to the data stored in the measurement data storage unit 40, calculates a probability density distribution (FIG. 14) at T0 + Δt after a predetermined time from the initial time using a memory-based particle filter, and the calculated value is calculated. A predicted value D (T0 + Δt) at T0 + Δt after a predetermined time from the initial time is calculated based on the obtained probability density distribution.

  More preferably, measurement data calculation unit 50 stores data D (T0) exceeding threshold value Yth only when data D (T0) obtained from measurement data storage unit 40 exceeds threshold value Yth. ) Is set as an initial value, and the prediction calculation is executed with the time T0 at which the data D (T0) exceeding the threshold Yth is measured as the initial time T0.

  More preferably, the measurement data calculation unit 50 obtains a similarity between the initial value D (T0) and each data stored in the measurement data storage unit 40, searches for a time Tref having a high similarity, and obtains the similarity. The time series data D (T0 + Δt) at a time T0 + Δt after a predetermined time from the initial time T0 is referred to by referring to the data D (Tref + Δt) at a time Tref + Δt after the predetermined time.

  According to the state monitoring device of the present embodiment described above, for example, when the acquired data exceeding the threshold value is detected, the data can temporarily rise even if there is no actual value after that time. Whether the probability is high can be notified together with a predicted value after a predetermined time from the initial time.

  For example, when spike noise candidate data is detected, a diagnosis can be performed based on a prediction result after a predetermined time from the initial time even if there is no actual value after that time.

  The embodiments disclosed this time are to be considered in all respects as illustrative and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description of the embodiments, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

  1 state monitoring device, 10 wind power generation device, 20 measuring device, 30 measurement data collecting unit, 40 measurement data storage unit, 50 measurement data calculation unit, 60 diagnostic unit.

Claims (6)

A measuring device,
A measurement data collection unit that collects data from the measurement device,
A measurement data storage unit that stores data obtained from the measurement data collection unit,
A measurement data calculation unit that calculates data obtained from the measurement data storage unit,
The measurement data calculation unit,
When the data obtained from the measurement data storage unit does not exceed a threshold set based on the data stored in the measurement data storage unit, the prediction calculation of the time-series data is not performed,
When the data obtained from the measurement data storage section exceeds the threshold value, the state monitoring device performs a prediction calculation of the time-series data using a time-series prediction model. Further comprising a diagnostic unit for diagnosing the state of the monitoring target based on the data collected by the measurement data collection unit,
The state monitoring device according to claim 1, wherein the diagnosis unit switches processing contents based on a result of a prediction calculation executed by the measurement data calculation unit. The time series prediction model is a memory-based particle filter,
The measurement data calculation unit refers to the data stored in the measurement data storage unit, calculates a probability density distribution after a predetermined time from an initial time using the memory-based particle filter, and calculates the calculated probability density distribution. The state monitoring device according to claim 1, wherein a predicted value after the predetermined time from the initial time is calculated based on the initial time. The data exceeding the threshold value as the initial value, the time when the data exceeding the threshold value is measured as the initial time,
The measurement data calculation unit executes the prediction calculation from the initial value and the initial time only when the data obtained from the measurement data storage unit exceeds the threshold value, according to claim 3, wherein State monitoring device as described. The measurement data calculation unit obtains a similarity between the initial value and each data stored in the measurement data storage unit,
The measurement data calculation unit searches for a time having a high degree of similarity and predicts time-series data after the predetermined time from the initial time by referring to data after the predetermined time from the time obtained by the search. The state monitoring device according to claim 4, which performs the operation.   A wind turbine generator comprising the condition monitoring device according to claim 1.