JP2010122912A - Abnormality decision device and method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To immediately perform state detection even under calculation resource restriction, to obtain a more accurate result than a threshold value method, and to clearly show grounds of abnormality detection. <P>SOLUTION: An abnormality decision device has: a noise filtering part for inputting sensor data and removing noise from the sensor data; a segmentation part for extracting partial waveform data of a proper length section from the sensor data removed with the noise; a model construction part for constructing a decision model based on the partial waveform data; a decision part for deciding, by use of the decision model, whether or not abnormality is present in unknown data when the unknown data not performed with decision are input; and an output part for outputting an abnormality decision result about the unknown data. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an abnormality determination device, method, and program capable of capturing an abnormality sign of a measurement target from sensor data and estimating an abnormal part.

  In recent years, with the spread of sensor networks, sensor data analysis techniques are required in various scenes. However, there is a problem that signal processing for correctly determining the state of a measurement target from various and varied sensor data is more difficult than conventional signal processing. For example, the relative phase shift of different sensor signals must be considered. Even for sensor nodes and low-cost computers and devices with severe computational resource constraints, the amount of computation must be kept low because data must be processed immediately.

  A method of detecting an abnormality by setting an abnormality threshold for sensor data in advance and determining whether the threshold has been exceeded (threshold method) does not take processing time but has a risk of increasing false alarms. In addition, when the amplitude of sensor data frequently gathers around the threshold, it is difficult to provide a clear basis for the determination (see, for example, Patent Documents 1 to 3 below). For this problem, a method such as fuzzing around the threshold can be considered, but this method is also difficult to present to the user as a clear basis in an easily understandable manner.

  Regarding the introduction of the adaptive filter technique, since a desired signal must be input, there is a problem that an abnormality determination is biased.

Furthermore, for technologies such as neural networks that are used to predict time-series data and are generally said to have high prediction accuracy, the model is a black box and it is difficult to verify why such state detection was performed. There is a problem that the basis of detection cannot be shown.
Japanese Patent No. 3624546 JP 2007-64307 A Japanese Patent No. 3978052

  The present invention has been made in consideration of the above circumstances, and can immediately detect the state even under computational resource constraints, obtain a more accurate result than the threshold method, and further clarify the grounds for abnormality detection. An object of the present invention is to provide an abnormal state determination device, method, and program that can be shown.

  An abnormality determination device according to an aspect of the present invention includes a noise filtering unit that inputs sensor data and removes noise from the sensor data, and partial waveform data of an appropriate length section from the sensor data from which the noise has been removed. A segmentation unit for extraction; a model construction unit for constructing a judgment model based on the partial waveform data; and when unknown data that has not been judged is input, whether the unknown data is abnormal or not And an output unit that outputs an abnormality determination result for the unknown data.

  In the abnormality determination method according to another aspect of the present invention, the noise filtering unit inputs sensor data and removes noise from the sensor data, and the segmentation unit has an appropriate length from the sensor data from which the noise has been removed. A step of extracting partial waveform data of a section; a step of a model construction unit constructing a determination model based on the partial waveform data; and when unknown data that has not been determined is input, an abnormality is detected in the unknown data A step of determining whether or not there is a determination unit using the determination model; and a step of outputting an abnormality determination result for the unknown data by the output unit.

  Another aspect of the present invention is a noise filtering unit that inputs sensor data and removes noise from the sensor data, a segmentation unit that extracts partial waveform data of an appropriate length section from the sensor data from which the noise has been removed, A model building unit that builds a determination model based on the partial waveform data, and a determination unit that determines whether or not there is an abnormality in the unknown data when unknown data that has not been determined is input A program for causing a computer, an embedded device, or the like to function as an output unit that outputs an abnormality determination result for the unknown data.

  According to the present invention, the channel sensor data accumulated in the past and the information indicating the state of the inspection target are utilized, and the features necessary for the determination are cut out and structured as a narrowing rule. Thus, it is possible to provide an abnormal state determination device that can detect a state and obtain a more accurate result than the threshold method. The threshold method focused only on the instantaneous peak of the value, but in the present invention, all the surrounding waveform information of the instantaneous peak of the value in the time series data of multiple channels and their phase shifts are considered and integrated. Can be determined.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram illustrating a configuration of an apparatus for capturing an abnormal sign of a measurement target from sensor data and enabling estimation of an abnormal part according to an embodiment of the present invention. This apparatus includes a noise filtering unit 1 that receives sensor data and removes noise, a segmentation unit 2 that extracts input sensor data only in an appropriate length section, and a model construction unit that learns a determination model from partial waveform data 3, a determination model storage unit 4 that stores the built determination model, a determination unit 5 that determines unknown data when unknown data without a determination result is input, and waveform data and determination results And an output unit 6 that informs the outside by a display or sound.

  First, sensor time-series data and data indicating abnormal locations / predictive locations are input to the noise filtering unit 1 as training input data (hereinafter referred to as training data). Abnormal parts and predictive parts are entered in advance in the database by maintenance personnel. The noise filtering unit 1 uses a low-frequency filter (low-pass filter) used in general signal processing, a smoothing filter that smooths the envelope of the waveform, and the like, and hum noise mixed in sensor data and noise caused by baseline fluctuations Etc. are removed.

  Next, the segmentation unit 2 takes out the data from which noise has been removed by the noise filtering unit 1 with a predetermined time interval around the time of the abnormal part / prediction part, and predicts the prediction segment (simple model). And The extracted data corresponds to a partial waveform.

  In normal data, partial waveforms in the time interval before and after the abnormal part designated by all data with abnormality are extracted. When this process is performed on all training data, the width of the predictive partial waveform is determined.

  If the sum of the deviations is calculated with a predetermined window width, the accumulated amount of deviations strongly depends on the window width. Accordingly, in the present invention, the cumulative amount of deviation is calculated by expanding the width one by one before and after the center time of the abnormal part / predictive part, up to the maximum section width that satisfies the condition that it is possible to distinguish between abnormal and normal. Adopt an expansion method. The width of the predictive partial waveform is determined by the following procedure, for example.

  (Step 1) N training data are divided into an abnormal group (n) and a normal group (m). i = 1.

  (Step 2) When i = n, the process is stopped. In the case of i <n, the i-th waveform data is extracted from the waveform data of n abnormal groups, and the waveform data before and after a certain interval is set in two different buffers A and B, respectively, with the sign point as the sign center position. Store (when data is stored in A, the data is stored in reverse order so that the sign center is stored first). Let s = 1.

  (Step 3) Predictor segment candidate range B from buffer center to point advanced by (s * Δ) in the buffer B is waveform data of n−1 abnormal groups other than the abnormal group waveform data extracted in 2 The sum of the deviation of the values from the partial waveform is calculated as Δ1i, and Δ1i is normalized by n−1.

  (Step 4) Similarly, the sum δ2i of the difference between the predictive segment candidate range B calculated in step 3 and the waveform data of all normal groups is calculated, and δ2i is normalized gradually by m.

  (Step 5) The normalized δ1i and δ2i calculated in Step 3 are compared, and if δ1i <δ2i, s = s + 1 and the process returns to 3. Otherwise, the segment B reaches a point advanced by (s * Δ) from the sign center in the buffer B, and the sign range is defined as (s * Δ) from the sign center.

  (Step 6) q = 1.

  (Step 7) In buffer A, the segment segment A that has been advanced by (q * Δ) from the center of the predictor is the predictor segment candidate range A. Waveforms of n−1 abnormal groups other than the abnormal group waveform data extracted in 2 The sum of the difference between the data and the partial waveform is calculated as Δ3i, and Δ3i is normalized by n−1.

  (Step 8) Similarly, the sum δ4i of the value deviation between the predictive segment candidate range B calculated in 6 and the waveform data of all normal groups is calculated, and δ4i is normalized gradually by m.

  (Step 9) The normalized δ3i calculated in Step 6 is compared with δ4i, and if δ3i <δ4i, s = s + 1 is returned to 6. Otherwise, the portion up to the point where (q * Δ) returns from the center of the sign in buffer A is set as the sign segment range A, and the part of (q * Δ) from the sign center is set as the sign range and i = i + 1 and the process returns to 2. .

  As described above, when cutting out the input training waveform, using the double buffer, the data section in the front-rear direction is cut out independently, and even if the window width before and after the data becomes asymmetric, the section cut is calculated at high speed It is preferable to expand so as to be possible.

  Of course, the calculation may be simplified by equalizing the width before and after the sign center. In this case, the portion of buffer B from step 3 to step 5 is processed “both buffers A and B simultaneously”, step 6 to step 9 are omitted, and “i = i + 1” is set to 2 in step 5. You just need to add a "return" process.

In order to calculate the difference (distance) between the amplitude values of the two waveforms, the Euclidean distance for simply comparing the absolute values of the amplitude differences at the same time may be used. The definition of the distance between waveforms is generally given as follows for 1 ≦ p ≦ ∞.

  Especially when p = 1, the Euclidean distance is used. In the Euclidean distance, the distance between waveforms is calculated at the same time without considering the shift of the peak phase of the time series data.

  For example, as shown in FIG. 2, sections having different widths are extracted before and after the center of the abnormal part / predictive part, and this abnormal / predictive pattern is extracted using the predictive segment range calculated here.

In order to calculate the accumulated amount of deviation, the DTW distance may be taken in consideration of local peak phase deviation. The DTW distance is calculated by the following procedure. First, the warping path W of the waveform x and the waveform y is obtained. The warping path W calculates time shift correction information for nonlinearly minimizing a local time shift between any two partial waveforms.

  W must satisfy the following conditions.

[Constraint 1] Boundary conditions

[Constraint 2] Continuity condition

[Constraint 3] Monotonicity condition

Here, in the distance matrix between waveforms, among the paths satisfying these constraints, the shortest path is obtained by the method shown in the following formula, and the distance obtained by the shortest path is defined as the DTW distance.

Actually, using the following recursive formula, a path of a pair of i and j passing through the shortest cumulative distance γ (i, j) is obtained and is set as W.

  The set of the amplitude value vector of the waveform fragment and the start time and end time of the waveform fragment (this is referred to as a simple model) relating to the sign / abnormal portion obtained in this way is passed to the model construction unit 3. The method of extending to the maximum section width that satisfies the condition that can be distinguished from abnormal and normal as described above is to determine the width by the above-mentioned simple calculation that makes the width before and after the sign center equal. Therefore, it is possible to calculate a partial waveform pattern that is more flexible and easier for the user to understand as a basis.

  The model construction unit 3 structures and combines the predictive segments calculated by the segmentation unit 2, that is, waveform simple fragments, performs internal verification for the number of training data, and only at least one simple model with high determination accuracy Is stored in the determination model storage unit 4 as a determination model. The lower limit value of the determination accuracy is determined by the user in advance, for example, 95%, and only a simple model exceeding this value is determined as a determination model.

  Next, a simple determination model obtained from the simple model described above is structured and subjected to internal verification, and a complex determination model with high determination accuracy is further constructed. An example of this procedure is as follows, for example. However, it is assumed that there are L simple models (waveform fragments).

  (Step 1) One of the L simple models or a determination model candidate is generated by any combination. When the generation of all judgment model candidates is completed, the process is stopped, and all judgment models in the judgment model set calculated so far are returned as results. Set the lower limit of the judgment probability.

  (Step 2) Using the generated determination model, training data whose determination result is known in advance is determined.

  (Step 3) The determination success / failure probability of Step 2 is calculated for the determination model candidate generated in Step 1, and a determination model candidate that exceeds the lower limit of the determination probability set in Step 1 is added to the determination model set.

  (Step 4) Return to Step 1.

  Here, the determination success / failure probability in step 3 may be accuracy, or a probability that a normal one is determined to be abnormal (false alarm probability) or a probability that an abnormal one is determined to be normal (missing probability) may be used.

  When the determination model in the determination model set is converted into the rule format, the expression becomes easy for the device user to understand (see FIG. 3). The basis of calculation of the determination model can be clarified by displaying the abnormality / predictive pattern on the display or the like in the output unit 6 in such a format.

  This is because, when an unknown input waveform is input in multiple channels, the similar input pattern near the occurrence time of the predictor segment is detected using a learned determination model. This indicates that an abnormal section can be estimated. Calculate the sum of the distance between the partial waveform of this predictive segment and the partial waveform of the same waveform in the waveform to be inspected, and calculate whether it is similar to the judgment model or normal data by the k nearest neighbor method Thus, it is possible to determine whether or not an abnormality occurs (FIG. 4). This is an indispensable determination method particularly in an apparatus configuration that determines whether there is an abnormality in the proximity of the distance between the waveform fragment of the past data and a part of the waveform of the inspection target data. In the k nearest neighbor method, at least one determination model for determining normality (normal determination model) and at least one determination model for determining abnormality (abnormality determination model) are respectively obtained from the head of the inspection target waveform. The time position is shifted and collated, and the distance between the waveform to be inspected at the time position where the distance is closest to the section where the predictive segment exists in the determination model is obtained, and the normal determination model and the abnormality determination model are in order of increasing distance. Of the top k models when the models are arranged, it is determined whether the number is determined to be normal or the number is determined to be normal, or whether the number is determined to be normal or abnormal. For example, when k = 3, when there are 100 normal determination models and 80 abnormality determination models, the sum of the distances between the corresponding portion of the waveform to be inspected and each point of the predictive pattern included in the model is calculated. As a result, if there is one case that is determined to be normal and two cases that are determined to be abnormal among the top three items having the smallest distance, it is determined (estimated) as abnormal by majority vote.

  When processing multi-channel waveforms simultaneously, the above method needs to be extended to multivariate. An embodiment in which multivariate waveforms are processed simultaneously is shown in FIG. The learned determination rule, that is, the determination model (abnormal sign pattern) has a plurality of channels as shown in FIG. However, channel 1 represents the x coordinate of the acceleration sensor, channel 2 represents the y coordinate of the acceleration sensor, and channel 3 represents the state of the detection switch that detects the ON or OFF state. An example of the abnormal range estimation in the determination unit 5 is shown in FIG. It can be seen that in the multivariate case, the predictive segment may be a heterogeneous sensor.

  Here, when the training input data is newly input, the model update unit 7 can dynamically update the determination model. The procedure is shown in the flowchart in FIG. First, newly input training data and parameters indicating the time of the abnormal range are input (step S1), and noise filtering is performed (step S2). Next, a newly input training waveform is dynamically segmented by the above method to generate a simple model (step S3). If this training data can be correctly determined by the determination model so far, the determination model is maintained without updating the model (step S4). If this training data cannot be determined correctly, a new simple model is generated, classified and verified using the training data, and a simple model exceeding the lower limit of accuracy is added to the determination model. Further, a simple model exceeding the lower limit of accuracy is structured, and a more complicated determination model is added to the determination model storage unit 4 (step S5). If a determination model stored in the determination model storage unit 4 has an adverse effect on the determination, the determination model may be partially deleted. Such processing is called model update.

  The abnormality determination result can also be applied to remote monitoring via the network unit 8 shown in the block diagram of FIG. FIG. 10 shows a more specific configuration of the apparatus that realizes such an embodiment. The data of various multi-channel sensors are synchronously input from the I / O via a multiplexer or the like, and the above abnormal state determination model is learned and unknown data is determined using the CPU, RAM, and ROM. When handling multi-channel data, the input / output data is the same as that shown in FIG.

  According to the embodiment of the present invention described above, it is possible to appropriately and simply detect the state of the measurement target object or the measurement target person from the channel sensor data that changes every moment. In addition, since the judgment model is composed of explicit predictive patterns, it is possible to show the reason why such state detection was performed. Maintenance sensor data in quality control / maintenance departments, biosensor data in medical care, etc. It is possible to detect and determine abnormalities and signs from time-series recordings of various events.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

1 is a block diagram of an apparatus for capturing an abnormality sign of a measurement target from sensor data and enabling estimation of an abnormal part according to an embodiment of the present invention. Diagram showing how to dynamically determine the waveform segment width Diagram showing learned abnormal / predictive patterns Schematic diagram of abnormality determination Diagram showing the case of processing multivariate waveforms simultaneously The figure which shows an example of a judgment rule (learned abnormal sign pattern) Diagram showing an example of abnormal range estimation Flow chart showing model update procedure Diagram showing the configuration for remote monitoring via a network Block diagram showing an example of a more specific hardware configuration

Explanation of symbols

1 ... noise filtering unit;
2 ... Segmentation part;
3 ... Model building department;
4 ... judgment model storage unit;
5 ... determination part;
6 ... output part;
7 ... Model update section

Claims (8)

A noise filtering unit for inputting sensor data and removing noise from the sensor data;
From the sensor data from which the noise has been removed, data sections in the front-rear direction are cut out independently using at least two buffers, and partial waveform data in an appropriate length section is extracted with the window width before and after the data asymmetry. A segmentation section;
A model building unit that builds a determination model based on the partial waveform data;
When unknown data that has not been determined is input, a determination unit that determines whether the unknown data has an abnormality using the determination model;
An output unit for outputting an abnormality determination result for the unknown data;
An abnormality determination device comprising:   The apparatus according to claim 1, further comprising a network unit that transmits and receives an abnormality determination result for the unknown data to another network node, a gateway, and the Internet / intranet via a network.   The apparatus according to claim 1, further comprising a model update unit configured to update the determination model when additional training data is input in a state where the determination model has already been constructed.   The apparatus according to claim 1, wherein the determination unit determines whether there is an abnormality based on a proximity of a distance between a waveform fragment of past data and a part of a waveform of inspection target data.   The apparatus according to claim 1, wherein the model construction unit generates the determination model for determining abnormality by integrating a group of waveform fragments as a sign and their time difference.   The apparatus according to claim 1, wherein the sensor data is a plurality of sensor data input simultaneously from a multi-channel sensor. A noise filtering unit that inputs sensor data and removing noise from the sensor data;
From the sensor data from which the noise has been removed, the segmentation unit independently cuts out the data section in the front-rear direction using at least two buffers, and obtains partial waveform data in an appropriate length section with the window width before and after the data asymmetry. Extracting, and
A model construction unit constructing a determination model based on the partial waveform data;
When unknown data that has not been determined is input, the determination unit determines whether the unknown data has an abnormality using the determination model; and
An output unit that outputs an abnormality determination result for the unknown data;
An abnormality determination method comprising: Computer
A noise filtering unit for inputting sensor data and removing noise from the sensor data;
From the sensor data from which the noise has been removed, data sections in the front-rear direction are cut out independently using at least two buffers, and partial waveform data in an appropriate length section is extracted with the window width before and after the data asymmetry. Segmentation section,
A model construction unit for constructing a determination model based on the partial waveform data;
When unknown data that has not been determined is input, a determination unit that determines whether the unknown data has an abnormality using the determination model,
An output unit for outputting an abnormality determination result for the unknown data;
Program to function as.

Images