An abnormal interval prediction model construction and abnormal interval prediction method

Abstract

The present invention provides an abnormal interval prediction model construction and an abnormal interval prediction method, which relate to the technical field of wind turbine status monitoring. The construction method of the present invention includes: obtaining wind turbine power generation data, annotating the wind turbine power generation data through an adaptive abnormal point annotation algorithm to obtain annotated data, constructing abnormal interval data according to the annotated data, and constructing an abnormal interval prediction model based on a multi-scale time series Transformer according to the abnormal interval data. The wind turbine power generation data is annotated by an adaptive abnormal point annotation algorithm to annotate abnormal points with trend changes, which effectively reduces the annotation cost and improves the annotation efficiency and accuracy. An abnormal interval prediction model based on a multi-scale time series Transformer is constructed based on the abnormal interval data, making full use of multi-scale time series information for prediction, and effectively improving the abnormal interval prediction accuracy, so as to realize the abnormal interval prediction of time series data, especially wind turbine power generation data.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine status monitoring, and in particular to an abnormal interval prediction model construction and an abnormal interval prediction method. Background Art

[0002] Time series is a sequence of sampled values ​​of a physical quantity of an objective object at different times arranged in chronological order. It is widely used in economic management and engineering. By using time series data mining, we can obtain useful time-related information contained in the data and achieve the purpose of extracting knowledge.

[0003] Anomalies in time series data refer to data patterns that have different data characteristics from normal situations. In various application fields, the occurrence of anomalies often provides key and actionable information. Therefore, by detecting the abnormal points of time series data, it is possible to provide guidance and improvement for actual design and application. However, only anomaly detection often cannot meet the actual application needs. Taking wind turbines as an example, when the abnormal situation of wind turbine power changes is detected by anomaly detection algorithms, the wind turbine power generation has often increased or decreased suddenly. At this time, there is a certain time lag in taking measures, and certain losses have already occurred. Therefore, only anomaly detection often cannot meet the actual application needs. Summary of the invention

[0004] The problem solved by the present invention is how to realize abnormal interval prediction of time series data, especially wind turbine power generation data.

[0005] To solve the above problems, the present invention provides a method for constructing an abnormal interval prediction model, including: acquiring wind turbine power generation data, labeling the wind turbine power generation data through an adaptive abnormal point labeling algorithm to obtain labeled data, constructing abnormal interval data according to the labeled data, and constructing an abnormal interval prediction model based on a multi-scale time series Transformer according to the abnormal interval data.

[0006] The abnormal interval prediction model construction method described in the present invention marks the wind turbine power generation data through an adaptive abnormal point marking algorithm to mark the abnormal points where trend changes occur, effectively reducing the marking cost and improving the marking efficiency and accuracy. An abnormal interval prediction model based on a multi-scale time series Transformer is constructed according to the abnormal interval data, making full use of multi-scale time series information for prediction, effectively improving the abnormal interval prediction accuracy, thereby realizing the abnormal interval prediction of time series data, especially wind turbine power generation data.

[0007] Optionally, the labeling of the wind turbine power generation data by an adaptive outlier labeling algorithm to obtain labeled data includes: obtaining a time window of a first preset length based on the wind turbine power generation data, and judging whether the data point is a normal point based on the time window; obtaining a long-scale time window of a second preset length, determining the standard deviation of the data in the long-scale time window, and initializing a set value change threshold; determining a corresponding threshold based on the traversal of the data point, and judging whether the data point is a normal point based on the threshold.

[0008] The abnormal interval prediction model construction method described in the present invention labels the wind turbine power generation data through an adaptive abnormal point labeling algorithm to label the abnormal points where trend changes occur, effectively reducing the labeling cost and improving the labeling efficiency and accuracy.

[0009] Optionally, obtaining a time window of a preset length based on the wind turbine power generation data includes: obtaining a data index, obtaining a corresponding data point in the wind turbine power generation data according to the data index, and obtaining the time window of the first preset length with the data point as the starting point.

[0010] The abnormal interval prediction model construction method described in the present invention obtains the corresponding data points in the wind turbine power generation data according to the data index to obtain the time window of the first preset length, thereby realizing adaptive abnormal point marking.

[0011] Optionally, judging whether the data point is a normal point according to the time window includes: judging whether the data point is the extreme value in the time window; if not, marking the data point as a normal point; if so, marking the extreme point of the opposite type to the data point in the time window as the extreme point.

[0012] The abnormal interval prediction model construction method described in the present invention determines whether a data point is a normal point according to a time window, thereby enabling adaptive abnormal point labeling.

[0013] Optionally, determining the corresponding threshold value according to the traversal of the data point includes: traversing the first preset length of steps to the right from the data point until encountering the maximum point or a change point of the same type as the data point; if the maximum point is encountered first, updating the set value change threshold value according to the standard deviation; if the change point is encountered first, determining the gradient threshold value according to the gradient growth rate.

[0014] The abnormal interval prediction model construction method described in the present invention determines the corresponding threshold value according to the traversal situation of the data point, so that it can judge whether the data point is a normal point according to the threshold value, and realize adaptive abnormal point labeling.

[0015] Optionally, judging whether the data point is a normal point according to the threshold value includes: determining a first comparison value according to a difference between a data array corresponding to the change point and a data array corresponding to the data point, determining a second comparison value according to a difference between a data array corresponding to the maximum point and a data array corresponding to the data point, and judging whether the data point is a normal point according to a comparison result of the first comparison value and the gradient threshold value, and a comparison result of the second comparison value and the set value change threshold value.

[0016] The abnormal interval prediction model construction method described in the present invention determines whether a data point is a normal point based on the comparison result between the comparison value and the threshold value, thereby realizing adaptive abnormal point labeling.

[0017] Optionally, constructing the abnormal interval prediction model based on the abnormal interval data includes: extracting time series feature information of different scales in the abnormal interval data through a time series feature extraction model based on Transformer-Encoder; fusing the time series feature information of different scales through an Attention-based classification model, using the fusion result as a classifier input to obtain a classification result, and constructing the abnormal interval prediction model based on the classification result.

[0018] The abnormal interval prediction model construction method described in the present invention adopts a Transformer-Encoder structure and fully utilizes multi-scale time series information for prediction based on Attention, thereby effectively improving the accuracy of abnormal interval prediction.

[0019] Optionally, the fusing of the temporal feature information of different scales through the Attention-based classification model includes: using classification tokens to fuse the temporal feature information of each time step, and weightedly fusing the long-term, medium-term and short-term temporal feature information based on the attention mechanism as classifier input to obtain a classification result.

[0020] The abnormal interval prediction model construction method described in the present invention adopts classification token to fuse the time series feature information of each time step, and based on the attention mechanism, weightedly fuses the long-term, medium-term and short-term time series feature information as the classifier input to obtain the classification result, fully utilizes multi-scale time series information for prediction, and effectively improves the prediction accuracy of the abnormal interval.

[0021] The present invention also provides an abnormal interval prediction method, comprising: obtaining wind turbine power data, and inputting the wind turbine power data into an abnormal interval prediction model constructed by the abnormal interval prediction model construction method to achieve abnormal interval prediction. The abnormal interval prediction method has the same advantages as the above-mentioned abnormal interval prediction model construction method over the prior art, which will not be repeated here.

[0022] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is read and executed by a processor, the above abnormal interval prediction model construction method or abnormal interval prediction method is implemented. The advantages of the computer-readable storage medium and the above abnormal interval prediction model construction method over the prior art are the same, which will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of a process for constructing an abnormal interval prediction model according to an embodiment of the present invention;

[0024] Figure 2 Schematic diagram of the structure of an abnormal interval prediction model based on a multi-scale time series Transformer according to an embodiment of the present invention;

[0025] Figure 3 2 is a structural diagram of a Transformer-Encoder according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0027] like Figure 1 As shown, an embodiment of the present invention provides a method for constructing an abnormal interval prediction model, including: acquiring wind turbine power generation data, labeling the wind turbine power generation data through an adaptive abnormal point labeling algorithm to obtain labeled data, constructing abnormal interval data according to the labeled data, and constructing an abnormal interval prediction model based on a multi-scale time series Transformer according to the abnormal interval data.

[0028] Specifically, in this embodiment, the abnormal interval prediction model construction method includes:

[0029] Obtain wind turbine power generation data, for example, obtain a wind turbine power generation data set from a wind turbine manufacturer (including status data and environmental data collected by sensors during wind turbine operation, such as instantaneous wind speed, ten-minute average wind speed, wind direction angle, nacelle position, rotor speed, active power generation, etc.), and preprocess the data, specifically including: Since the wind turbine data is collected once every half a minute, which is not aligned with the actual prediction step (ten minutes), the half-minute data is first aggregated and converted into ten-minute data; secondly, since the power generation curve is very steep, which is not conducive to model prediction, a wavelet transform is performed on it to smooth the data.

[0030] Among them, this embodiment realizes the denoising function by using wavelet transform, and the specific steps include: firstly, selecting a suitable orthogonal wavelet basis as the decomposition basis, and then selecting the number of layers k to decompose the time series; using wavelet transform to decompose the time series containing noise into k layers to obtain k groups of wavelet coefficients; performing threshold processing on the decomposed wavelet coefficients, setting the wavelet coefficients with smaller values ​​to zero, and reducing the larger wavelet coefficients to a certain extent to make them close to zero; using the wavelet coefficients after threshold processing to reconstruct the time series through inverse wavelet transform to obtain the denoised time series.

[0031] The wind turbine power generation data is labeled by an adaptive outlier labeling algorithm to obtain labeled data, in order to mark the outliers with trend changes. The core idea is to use the sliding window method, take the standard deviation of the data in the time window as the basic division threshold, and refine the division thresholds for rapidly changing outliers and progressively changing outliers.

[0032] After labeling, the data set is divided into a training set and a test set. Then the data in the training set is divided into small patches (batches) and sent to the constructed deep neural network for training. After the training is completed, the data in the test set is sent to the network to evaluate the results of the network training. The data in the test set is complete and does not need to be segmented. After that, the daily data of the wind turbine is continuously retrieved through the data interface and sent to the model for abnormal prediction. At the same time, the data of the previous day will be sent to the model for incremental training to enhance the performance of the model.

[0033] According to the labeled data, abnormal interval data based on multi-scale time series Transformer is constructed. A column of interval abnormal point attributes is added to the original data set, and then calculation starts from a certain time point Pi (i = 1, 2, 3, 4...n). If there is at least one abnormal point in an interval D with a length of width, then the i-th sample of the marked data set has an interval abnormal point.

[0034] An abnormal interval prediction model is constructed according to the abnormal interval data. The abnormal interval prediction model includes two parts, the first part is a time series feature extraction model based on Transformer-Encoder, and the second part is a classification model based on Attention.

[0035] In this embodiment, the wind turbine power generation data is labeled by an adaptive outlier point labeling algorithm to label the outliers where trend changes occur, which effectively reduces the labeling cost and improves the labeling efficiency and accuracy. An outlier interval prediction model based on a multi-scale time series Transformer is constructed according to the outlier interval data, which makes full use of multi-scale time series information for prediction and effectively improves the accuracy of outlier interval prediction, thereby enabling the prediction of outlier intervals of time series data, especially wind turbine power generation data.

[0036] Optionally, the labeling of the wind turbine power generation data by an adaptive outlier labeling algorithm to obtain labeled data includes: obtaining a time window of a first preset length based on the wind turbine power generation data, and judging whether the data point is a normal point based on the time window; obtaining a long-scale time window of a second preset length, determining the standard deviation of the data in the long-scale time window, and initializing a set value change threshold; determining a corresponding threshold based on the traversal of the data point, and judging whether the data point is a normal point based on the threshold.

[0037] Specifically, in this embodiment, the algorithm parameters of the adaptive outlier marking algorithm include the gradient growth rate r, the maximum growth coefficient scale, the time window length width, the window observation scale K, the data array V, and the data index Index, wherein r and scale determine the division threshold of the outlier, width and K determine the time window size observed by the data point, the data array V contains the actual wind turbine power generation data, and the data index contains the subscript corresponding to the data in the array. The specific steps of the algorithm include:

[0038] (1) Get the current data point P i As the starting point, the time window D with a length of width is used to determine whether the current point is the maximum value in the time window. If not, it is marked as a normal point. Otherwise, it is recorded as the maximum value in the time window. i The maximum point of the opposite type is point M, which is at a distance of P i The distance D M , go to step 2.

[0039] (2) Get the P i A long-scale time window with the point as the center and the window length of 2K*width is used to calculate the standard deviation σ of the data in the window, and the initial setting value change threshold β=scale, α=0, and go to step 3.

[0040] (3) From P i Start from point , and traverse width steps to the right until you encounter the maximum value M or the value that matches P i The same type of change point P S :

[0041] If the maximum point M is encountered first, then set the value change threshold β = scale * σ and go to step 4;

[0042] If you first encounter P S , and its distance from the starting point is D S , then calculate the gradient threshold Go to step 4.

[0043] (4) If and Then the data at the i-th position of the array Label is marked as an abnormal point, otherwise it is marked as a normal point.

[0044] In the time interval with large data fluctuations, there is a relatively large anomaly classification threshold to screen out the more important trend anomalies and filter out the secondary trend anomalies in the middle. In the place where the data fluctuations are small, there is a relatively small classification threshold to find as many potential anomalies as possible.

[0045] In this embodiment, the wind turbine power generation data is annotated by using an adaptive outlier point annotation algorithm to annotate outliers where trend changes occur, thereby effectively reducing the annotation cost and improving the annotation efficiency and accuracy.

[0046] Optionally, obtaining a time window of a preset length based on the wind turbine power generation data includes: obtaining a data index, obtaining a corresponding data point in the wind turbine power generation data according to the data index, and obtaining the time window of the first preset length with the data point as the starting point.

[0047] Specifically, in this embodiment, each index i in the index array Index is traversed to obtain its corresponding data point P in the wind power data V. i , then in V with the current data point P i As the starting point, get the time window D with a length of width.

[0048] In this embodiment, the time window of the first preset length is obtained by acquiring corresponding data points in the wind turbine power generation data according to the data index, so as to realize adaptive abnormal point marking.

[0049] Optionally, judging whether the data point is a normal point according to the time window includes: judging whether the data point is the extreme value in the time window; if not, marking the data point as a normal point; if so, marking the extreme point of the opposite type to the data point in the time window as the extreme point.

[0050] Specifically, in this embodiment, judging whether a data point is a normal point according to the time window includes: judging whether the current point is the maximum value in the time window, if not, marking the data as a normal point at the i-th position of the array Label, otherwise marking the data as the maximum value in the time window. i The maximum point of the opposite type is point M, which is at a distance of P i The distance D M .

[0051] In this embodiment, whether a data point is a normal point is determined according to the time window, so that adaptive outlier marking can be achieved.

[0052] Optionally, determining the corresponding threshold value according to the traversal of the data point includes: traversing the first preset length of steps to the right from the data point until encountering the maximum point or a change point of the same type as the data point; if the maximum point is encountered first, updating the set value change threshold value according to the standard deviation; if the change point is encountered first, determining the gradient threshold value according to the gradient growth rate.

[0053] Specifically, in this embodiment, determining the corresponding threshold according to the traversal of the data point includes: i Start from point , and traverse width steps to the right until you encounter the maximum value M or the value that matches P i The same type of change point P S If the maximum value point is encountered first, the set value change threshold is updated according to the standard deviation. If the change point is encountered first, the gradient threshold is determined according to the gradient growth rate.

[0054] In this embodiment, the corresponding threshold is determined according to the traversal situation of the data point, so that it is possible to judge whether the data point is a normal point according to the threshold, thereby realizing adaptive outlier point marking.

[0055] Optionally, judging whether the data point is a normal point according to the threshold value includes: determining a first comparison value according to a difference between a data array corresponding to the change point and a data array corresponding to the data point, determining a second comparison value according to a difference between a data array corresponding to the maximum point and a data array corresponding to the data point, and judging whether the data point is a normal point according to a comparison result of the first comparison value and the gradient threshold value, and a comparison result of the second comparison value and the set value change threshold value.

[0056] Specifically, in this embodiment, judging whether a data point is a normal point according to the threshold value includes: determining a first comparison value according to the difference between the data array corresponding to the change point and the data array corresponding to the data point. Determine the second comparison value based on the difference between the data array corresponding to the maximum point and the data array corresponding to the data point according to and The comparison result determines whether the data point is a normal point.

[0057] In this embodiment, whether a data point is a normal point is determined based on the comparison result between the comparison value and the threshold value, thereby realizing adaptive outlier point marking.

[0058] Optionally, constructing the abnormal interval prediction model based on the abnormal interval data includes: extracting time series feature information of different scales in the abnormal interval data through a time series feature extraction model based on Transformer-Encoder; fusing the time series feature information of different scales through an Attention-based classification model, using the fusion result as a classifier input to obtain a classification result, and constructing the abnormal interval prediction model based on the classification result.

[0059] Specifically, in this embodiment, combined with Figure 2 As shown in Figure 2, the abnormal interval prediction model consists of two parts. The first part is a time series feature extraction model based on Transformer-Encoder, and the second part is a classification model based on Attention. Figure 3 The Transformer-Encoder model structure shown in the figure (including the multi-head attention mechanism Multi-Headattention and the feedforward neural network, etc.) initially passes through an input embedding layer for dimension upgrading, and then is input into the Encoder after adding the position encoding, and finally the extracted time series feature information is obtained.

[0060] The Transformer-Encoder model introduces the classification token in the NLP classification task to aggregate the representation information of the entire sequence to facilitate the subsequent classification task. This token without obvious information will more fairly integrate the feature information of each time step. Specifically, Self-attention is a method of using the information of other time steps in the sequence to enhance the feature representation of the current time step, but the information of the current time step itself still accounts for the majority. Therefore, after multiple layers of Encoder Block, the information of each time step integrates the information of all time steps and can better represent its own information. The CLS bit itself has no feature information. After multiple layers of Encoder Block, what is obtained is the weighted average of all time steps, which can better represent the global feature information compared to other time steps.

[0061] The encoder consists of multiple identical layers, each of which consists of two sublayers, namely the multi-head attention layer and the fully connected feedforward network layer. Each sublayer has added residual connections and regularization to effectively avoid the effects of gradient vanishing or gradient explosion and model degradation. The fully connected feedforward network layer is mainly responsible for providing nonlinear transformations, while the multi-head attention layer projects Q, K, and V through h different linear transformations, and finally concatenates the different attention results:

[0062] MultiHead(Q,K,V)=Concat(head1,…,head h )W U

[0063]

[0064] Attention is calculated by scaling the dot product:

[0065]

[0066] The parameter Q represents the query matrix, K represents the key-value mapping matrix, V represents the input feature matrix, and d k is the scaling factor.

[0067] Since it is difficult to capture the temporal features of a time series using a single Transformer-Encoder, the time series can be split into three segments: short-term, medium-term, and long-term. Taking a time series of length 12 as an example, it is split into three time series of lengths 4, 8, and 12, which are input into the Transformer-Encoder model respectively to obtain the three-segment temporal features. The features are then integrated through the attention mechanism and finally input into the classification layer to obtain the classification result. The classification result is determined based on the predicted probability. If the probability of the prediction being abnormal is greater than the prediction being normal, it is determined as an abnormal interval, otherwise it is determined as a normal interval.

[0068] In this embodiment, a Transformer-Encoder structure is adopted, and multi-scale time series information is fully utilized for prediction based on the Attention method, which effectively improves the prediction accuracy of abnormal intervals.

[0069] Optionally, the fusing of the temporal feature information of different scales through the Attention-based classification model includes: using classification tokens to fuse the temporal feature information of each time step, and weightedly fusing the long-term, medium-term and short-term temporal feature information based on the attention mechanism as classifier input to obtain a classification result.

[0070] Specifically, in this embodiment, fusing temporal feature information of different scales through an Attention-based classification model includes: using classification tokens to fuse the temporal feature information of each time step, and weightedly fusing long-term, medium-term and short-term temporal feature information based on the attention mechanism as classifier input to obtain a classification result.

[0071] In this embodiment, classification tokens are used to fuse the time series feature information of each time step, and the long-term, medium-term and short-term time series feature information are weighted and fused as the classifier input based on the attention mechanism to obtain the classification result, making full use of multi-scale time series information for prediction, and effectively improving the prediction accuracy of abnormal intervals.

[0072] Another embodiment of the present invention provides an abnormal interval prediction method, comprising: acquiring wind turbine power generation data, and inputting the wind turbine power generation data into an abnormal interval prediction model constructed by the abnormal interval prediction model construction method to achieve abnormal interval prediction.

[0073] Another embodiment of the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is read and executed by a processor, the above-mentioned abnormal interval prediction model construction method or abnormal interval prediction method is implemented.

[0074] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A method for constructing an abnormal interval prediction model, characterized in that: include: Acquire wind turbine power generation data, annotate the wind turbine power generation data by using an adaptive anomaly point annotating algorithm to obtain annotated data, construct anomaly interval data according to the anomaly interval data, and construct an anomaly interval prediction model based on a multi-scale time series Transformer according to the anomaly interval data; The step of labeling the wind turbine power generation data by an adaptive outlier labeling algorithm to obtain the labeled data includes: Acquire a time window of a first preset length according to the wind turbine power generation data, and determine whether a data point is a normal point according to the time window; wherein, acquiring a time window of a first preset length according to the wind turbine power generation data comprises: acquiring a data index, acquiring a corresponding data point in the wind turbine power generation data according to the data index, and acquiring the time window of the first preset length with the data point as the starting point; Acquire a long-scale time window of a second preset length, determine a standard deviation of the data in the long-scale time window, and initialize a setting value change threshold; A corresponding threshold is determined according to the traversal condition of the data point, and whether the data point is a normal point is determined according to the threshold.

2. The abnormal interval prediction model construction method according to claim 1 is characterized in that: The determining whether the data point is a normal point according to the time window comprises: Determine whether the data point is the maximum value in the time window. If not, mark the data point as a normal point. If so, mark the maximum point in the time window that is opposite to the data point type as the maximum point.

3. The abnormal interval prediction model construction method according to claim 2 is characterized in that: Determining the corresponding threshold according to the traversal situation of the data point includes: Traverse the first preset length of steps to the right from the data point until the extreme point or a change point of the same type as the data point is encountered. If the extreme point is encountered first, update the set value change threshold according to the standard deviation. If the change point is encountered first, determine the gradient threshold according to the gradient growth rate.

4. The abnormal interval prediction model construction method according to claim 3 is characterized in that: The determining whether the data point is a normal point according to the threshold value comprises: A first comparison value is determined according to a difference between a data array corresponding to the change point and a data array corresponding to the data point, a second comparison value is determined according to a difference between a data array corresponding to the maximum point and a data array corresponding to the data point, and whether the data point is a normal point is determined according to a comparison result between the first comparison value and the gradient threshold and a comparison result between the second comparison value and the set value change threshold.

5. The abnormal interval prediction model construction method according to claim 1 is characterized in that: The constructing of an abnormal interval prediction model based on a multi-scale time series Transformer according to the abnormal interval data includes: Extracting time series feature information of different scales in the abnormal interval data through a time series feature extraction model based on Transformer-Encoder; The time series feature information of different scales is fused through an Attention-based classification model, the fusion result is used as a classifier input to obtain a classification result, and the abnormal interval prediction model is constructed according to the classification result.

6. The abnormal interval prediction model construction method according to claim 5 is characterized in that: The fusing of the temporal feature information of different scales through the Attention-based classification model includes: The classification token is used to fuse the temporal feature information of each time step, and the long-term, medium-term and short-term temporal feature information is weightedly fused based on the attention mechanism as the classifier input to obtain the classification result.

7. A method for predicting abnormal intervals, characterized in that: include: The wind turbine power generation data is obtained, and the wind turbine power generation data is input into the abnormal interval prediction model constructed by the abnormal interval prediction model construction method according to any one of claims 1 to 6 to realize abnormal interval prediction.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is read and executed by a processor, the abnormal interval prediction model construction method according to any one of claims 1 to 6 or the abnormal interval prediction method according to claim 7 is implemented.

Images